Master Thesis Multivariate Statistical Process Control in Industrial Plants Section: Energy & Industry (E&I) Student: Wenchang Chen (Vincent) Student no.: 1190229 Supervision committee: Prof. Dr. ir. M.P.C. Weijnen (E&I) Dr. ir. Zofia Verwater-Lukszo (E&I) Dr. Eric Molin (TLO) Dr. Alessandro Di Bucchianico (TUE) A.M.J. Sinon - B.Sc. (Sappi) Department of Engineering and Policy Analysis Faculty of Technology, Policy and Management Delft University of Technology, August 2005 Executive summary Operational management is a common issue; it addresses the problem that the academic research knowledge is not well implemented into practical filed. Multivariate statistical process control (MSPC) is one of these issues. Therefore, the research objective of this thesis research is to make recommendations for implementing multivariate statistical process control (MSPC) in a process-industry plant by providing clear interpretations of MSPC and suggestions to quality management staff in the plant. Being trained in the Faculty of Technology, Policy and Management, we will look at the technical problem with a broader view and provide promising solutions while considering several relevant aspects, such as finance, management, organization, etc. In this research, both academic development of MSPC and the status of its application in practical field will be investigated. By knowing exactly what the gap is between academic and practical fields, we will further propose practical tools and recommendations to fill up the gap – facilitate the implementation of MSPC in the industrial plants. Before discussing what MSPC is, some fundamental background knowledge of statistical process control (SPC) is necessary be introduced. The aim of SPC is to achieve higher quality of final product and lower the production loss due to the defect products. Process monitoring with control chart is a basic tool of SPC. Control chart monitors the behavior of a production process and signals the operator to take necessary action when abnormal event occurs. One of the most widely adopted control charts was developed by Dr. Shewhart (Shewhart, 1931); it is also called Shewhart control charts. Although the Shewahrt control chart is considered as simple and easy to understand, it monitors the variables separately and the relations between variables are ignored. Nowadays the production process has become more complex than it was in the past. Numerous variables need to be monitored, and they are often mutually correlated, which means a certain relationship existing between variables. Under such circumstance, assuming the variables are independent of each other can be insufficient on detecting process variation. Speaking of the relationship between variables, we come to the issue of multivariate statistical process control (MSPC). MSPC can be traced back to Hotelling’s T2 method (Hotelling, 1931). This method considers the correlation between variables, and monitors more than one variable simultaneously. By monitoring the relationship between variables, MSPC reflects the process situation more precisely and is able to detect the out-of-control event due to anti-correlation. Despite of these advantages, nevertheless MSPC has some drawbacks, for example, involving complex statistics, difficult to interpret the result from MSPC, losing systematic pattern, etc. These barriers indeed have weakened the chance of implementing this technique into practical field. On the other hand, what is the perspective from practical field on the issue of implementing MSPC? Various aspects were investigated, such as MSPC seems powerful, but really complex; are the quality people able to understand and conduct MSPC; will it be profitable to proceed the implementation and so on. These practical concerns indeed captured our attention during this research. Combining the perspectives from academic field and practical field provides us a clearer overview of the gap between them, so we would be able to develop practical tools to fill up the gap and support the industrial plants to enjoy the benefit by applying advanced technology. During this research, we developed MSPC Implementation Guideline, which contains four elements, MSPC Plan, MSPC Training, Team Approach and Management Involvement. Especially in MSPC Plan, two practical tools were constructed, which are Method Model of (M)SPC and MSPC Diagnosis. Method model of (M)SPC is a decision flow chart, which supports the practitioners to apply the proper statistical process control chart for different circumstances. It covers the situations of using Shewhart control charts and using MSPC control chart. MSPC Diagnosis is designed to interpret the result of MSPC. Because the MSPC only signals the occurrence of an out-of-control event; it does not provide further information about what the problematic variable(s) are. We developed these tools based on the existing theoretical methods while considering the acceptance of practitioners, for instance we tried to design the approach as simple as possible to increase the workability while maintaining the correctness and effectiveness. The other parts of MSPC Implementation Guideline emphasize how to practically implement this technique into industrial plants and the concerns of financial, organizational, managerial aspects were incorporated as well. The research development is validated with a case study. The case is a part of the paper-making process. We started with process investigation and tried to understand the mechanism of the process. After that, three variables were selected to apply Method Model of (M)SPC and MSPC Diagnosis. The analysis result validated the tools are effective and workable, even when the variables are not highly correlated in this case. It raises our confidence and the value of MSPC technique, because the probability that an out-of-control cannot be detected by Shewhart control charts but can be signaled by MSPC will become larger when the variables are highly correlated. However, by analyzing this case, we have also realized dealing with real problem is more difficult than analyzing the simulated data from scientific paper. Because the real process system is actually very complex and the simulated data has been often simplified to have a clear framework for explanation and demonstration. The process system in this case, several automatic control devices and loop control systems were included. Some variables change all the time due to automatic controller, and these variables are not suitable to be monitored by using statistical process control chart. Under such circumstance, we applied another technique – multiple regression analysis to analyze the relation between variables. All in all, several ideas generated from this case study. First, the process investigation is very important. Understanding the process correctly and monitoring the critical variables are fundamental of applying MSPC technique. Secondly, other SPC techniques are definitively required, for example design of experiment (DoE), multiple regression analysis and etc. Choosing the proper technique for different situations is crucial to achieve overall process performance improvement, and MSPC is one of these techniques. The entire research can be summarized as follows. We started with the literature study of MSPC. The nature of the MSPC and especially regarding the diagnosis of responsible variable(s) from MSPC result were discussed in detail. Then we turned to investigate the perspective from practitioners regarding the implementation of MSPC technique in practice. Implementing a particular technique is not simply a technical issue, in stead, financial, managerial, organizational aspects were involved as well. With a broader view of this issue, we developed the MSPC Implementation Guideline and applied it to a case study – a process unit from paper-making production. Several findings and recommendations, which can be a good reference for the process industrial plants, were generated during the entire period of research. Although advanced techniques are often more complex than the existing one, it is still possible to apply them in the practical field. It should be aware that the concerns of academic researchers and of practitioners may differ, and they are not simply technical issues. By investigating and understanding the gap with broader perception, we will be able to construct the link between both sides and support the practical field to utilize the benefit of advanced technology. Acknowledgement Although the so-called TBM thesis market was opened in the October of 2004, where the possible thesis research topics were presented to EPA program students, the seed of my decision already started to bud in the beginning of 2004. At that time, I took the elective courses of Integrated Plant Management, and Operation Analysis for Quality Management, and they did capture my attention. The manufacturing industries, the quality of production, using statistical technique to improve the process performance, how to lower the production cost ,how to increase the profit of company and so on motivate me to investment the last period of master study on this subject. This master thesis is a product from six-month process (without statistical process control) but it contains the contributions and efforts by numerous supportive experts. I would like to thank my supervision committee: Professor Weijnen, Dr. Verwater-Lukszo, Dr. Molin, Dr. Bucchianico and Mr. Sinon for their knowledgeable comments and instructions. Especially to Dr. Verwater-Lukszo, being my daily supervisor, her intensive dedication is highly appreciated. Besides, I would like to thank Mr. Telman, Mr. Mooiweer, and Mr. Proper for their valuable experience and knowledge. I would consider this thesis research period is a fantastic experience in my life. Not only academic enrichment, many tacit gains, such as cultural impact, cultivation of independent thinking, strengthening my confidence and so on, are very precious. And the most important thing is……. I do enjoy it!! 05/Aug/2005, Wenchang Chen (Vincent) Table of Contents Chapter 1. Introduction ..............................................................................................................1 1.1. Research Background ................................................................................................... 1 1.2. Research Questions and the Objective......................................................................... 2 Chapter 2. Statistical Process Control .....................................................................................5 2.1. Shewhart Control Charts .............................................................................................. 5 2.1.1. Control Limits ...................................................................................................... 6 2.1.2. Patterns of Process Behavior.............................................................................. 7 2.1.3. Control Charts for Attributes................................................................................ 9 2.1.4. Control Charts for Variables ................................................................................ 9 2.2. Multivariate Statistical Process Control ....................................................................... 10 2.2.1. Hotelling’s T2 Statistic........................................................................................... 12 2.2.2. T2A & SPE Plot .................................................................................................. 15 2.3. Diagnostic Approaches for Hotelling’s T2 Method ....................................................... 18 2.3.1. MYT T2 Decomposition ..................................................................................... 18 2.3.2. T2 Diagnosis with Principal Component Analysis (PCA)................................... 23 2.4. Approaches Discussion............................................................................................... 26 Chapter 3. Industrial Practice..................................................................................................29 3.1. Selection of the Interviewees ...................................................................................... 29 3.2. Objectives of the Interviews ........................................................................................ 30 3.3. Insights from the Interviews......................................................................................... 30 3.4. The Gap between Academic Field and Practical Field of Statistical Issues................ 32 3.5. Conclusions of the Interviews...................................................................................... 35 Chapter 4. MSPC Implementation Guideline..........................................................................38 4.1. MSPC Plan ................................................................................................................. 38 4.1.1. Method Model for (M)SPC ................................................................................ 38 4.1.2. MSPC Diagnosis ............................................................................................... 43 4.1.3. Process control ................................................................................................. 44 4.2. MSPC Training ........................................................................................................... 47 4.3. Team Approach .......................................................................................................... 47 4.4. Management Involvement .......................................................................................... 48 Chapter 5. Case Study..............................................................................................................49 5.1. Case Briefing ............................................................................................................... 49 5.2. MSPC Implementation ............................................................................................... 50 5.3. Result of MSPC Implementation ................................................................................. 57 5.4. Reflection..................................................................................................................... 60 5.5. Process Performance Improvement............................................................................ 61 Chapter 6. Conclusions and Recommendations...................................................................65 6.1. Conclusions ................................................................................................................. 65 6.2. Recommendations....................................................................................................... 68 6.3. Future Research Prospect .......................................................................................... 69 Reference .....................................................................................................................................70 Appendix A. Formulas of Shewhart Control Charts..............................................................72 Appendix B. Constants for Selected Control Charts ............................................................77 Appendix C. Hawkins’ Data Set and T2 Statistics of Measurements. ..................................78 Appendix D. Questionnaires for Interviews. ..........................................................................80 Appendix E. Multiple Regression Analysis of Sub-process. ...............................................85 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 1 List of Figures Figure 1-1 Structure of the thesis. ............................................................................................... 4 Figure 2-1 A generic Shewhart control chart. .............................................................................. 6 Figure 2-2 Typical systematic patterns. ....................................................................................... 8 Figure 2-3 The Western Electric run rules................................................................................... 8 Figure 2-4 An example of misleading information generated from Shewhart control chart. ......11 Figure 2-5 A generic bivariate Hotelling’s T2 control region....................................................... 12 Figure 2-6 A generic T2 control chart. ........................................................................................ 14 Figure 2-7 T2 control chart of measurements 36 to 50. ............................................................. 15 Figure 2-8 TA2 control chart based on first two principal components. ...................................... 17 Figure 2-9 SPE chart based on first two principal components................................................. 17 Figure 2-10 Normalized PCA scores. ........................................................................................ 24 Figure 2-11 Variable contribution plot of principal component 2................................................ 25 Figure 2-12 Variable contribution plot of principal component 3. .............................................. 25 Figure 2-13 Overall average contribution per variable. ............................................................. 26 Figure 2-14 A structure of multivariate statistical process control approaches. ........................ 27 Figure 4-1 MSPC Implementation guideline.............................................................................. 38 Figure 4-2 Method model for (M)SPC ....................................................................................... 42 Figure 4-3 MSPC diagnosis....................................................................................................... 44 Figure 4-4 An example of Causal-and-effect diagram............................................................... 45 Figure 4-5 The Shewhart/Deming wheel (PDCA)...................................................................... 46 Figure 5-1 Scheme of paper-making process unit..................................................................... 50 Figure 5-2 I-chart of HBtotal with one-minute measurement interval. ....................................... 51 Figure 5-3 Histogram of HBtotal measurements. ...................................................................... 52 Figure 5-4 I-chart of HBtotal with 20-minute interval. ................................................................ 52 Figure 5-5 MSPC decision path of case study........................................................................... 53 Figure 5-6 T2 control chart of in-control measurements. ........................................................... 55 Figure 5-7 T2 control chart for future observations. ................................................................... 56 Figure 5-8 Overall average contribution of every variable from observation 22 to 28. ............. 57 Figure 5-9 I-chart of Filler, HBtotal, HBash and BW. ................................................................. 59 Figure 5-10 Ellipse control chart of HBtotal and HBash. ........................................................... 60 Figure 5-11 Sub-process. .......................................................................................................... 61 Figure 5-12 Contour plot of HBtotal and Ret with respect to BW. ............................................. 63 List of Tables Table 2-1 Unique MYT decomposition terms. (cited from Mason, Young & Tracy, 1997) ......... 20 Table 2-2 Individual Ti2 and its status......................................................................................... 22 Table 2-3 Bivariate conditional term and its status. ................................................................... 22 Table 3-1 List of interviewees .................................................................................................... 29 Table 3-2 The gap between academic and practical fields........................................................ 32 Table 5-1 Correlation of process variables ................................................................................ 54 Table 5-2 In-control process data .............................................................................................. 56 Table 5-3 The result of MSPC diagnosis. .................................................................................. 58 Table 5-4 Result of regression model of sub-process. .............................................................. 62 Table 5-5 Possible recipe of HBtotal and Ret ............................................................................ 64 Chapter 1. Introduction Chapter 1. Introduction In this chapter, fundamental information of this research will be introduced, such as research background, research motivation, research scope, research questions and objectives. At the end of this chapter, an overview of the whole report will be provided. 1.1. Research Background There are various definitions of quality; one is that “Quality is the totality of features and characteristics of a product or service that bear on its ability to satisfy stated or implied needs” defined by International Organization for Standardization (ISO). In an easier expression, quality means to what extent the products can meet the requirements defined by the customers. High quality of product is the vital concern for most of the companies that will survive in this highly competitive global market. One of the most effective approaches to achieve high product quality is Statistical Process Control (SPC). Statistical Process Control (SPC) has become an important approach for process industries since 1920s. The aim of SPC is to achieve higher product quality and lower the production cost due to the minimization of the defect product. One of the greatest tools is the statistical process control chart developed by Dr. Walter A. Shewhart (Shewhart, 1931). He also pointed out an important fact that variation of a process is resulted from two sources. One is termed as common causes which are inherent in the production system and it is not possible to remove it, and the other is termed as special causes which are resulted from several particulate reasons (e.g. problems with raw material, operator mistakes, machine failures, etc.) and special causes may lead to serious damage to the product quality. In general, statistical process control techniques help us to monitor the production process and to detect abnormal process behavior due to special causes. The idea is very straightforward, once the special causes of abnormal process behavior can be detected and further eliminated; the process can be improved, so as the quality of product. However, an important characteristic of Shewhart control chart is that it can only monitor single process variable at a time. Nowadays, the modern production process has dramatically become complex and integrated. Monitoring the process variables separately ignores the possible correlation or interaction between them and thus Shewhart’s approach is criticized as inadequate to reflect the process situation sufficiently. For example (Kourti & MacGregor, 1996), in a high-pressure low-density polyethylene (LDPE) reactor, an increase of impurity in ethylene inhibits the polymerization process; moreover, fouling of the reactor walls by sticky polymer impedes heat transfer and cooling of the reactor. Both impurities and fouling cannot be measured directly. When these problems occur they affect several process variables and eventually the product quality. Their existence can therefore be detected by the effect they have on the process and quality variables. This effect is not a simple shift of mean of one or more variables; both the magnitudes and the relationships of the variables to each other will change. Using traditional Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 1 Chapter 1. Introduction Shewhart control chart is not capable to reflect such complex process behavior and to detect the problem of process. Therefore some improvement of Shewhart control chart was developed. One of the most often discussed methods is Hotelling’s T2 method (Hotelling, 1931). Hotelling’s T2 method considers the correlation between process variables, and it can generate control limits to monitor whether the process behavior is stable and detect variation resulted from special cause. Hotelling’s T2 method can simultaneously monitor more than one process variables at a time, and that is why it is also called multivariate statistical process control (MSPC) chart. Nevertheless, along with the advantage of MSPC, there are also several shortcomings that need to be mentioned. First, the result of MSPC compares the synthetic statistic value generated from more than one process variables with the calculated control limit generated from a period of in-control historical data set. MSCP can detect an abnormal event but does not provide a reason for it. It is difficult for a user to identify which process variable or set of process variables is responsible for the abnormal event and to take necessary action. Second, the application of MSPC involves too much statistic knowledge for plant staff in the industries. Due to the complex nature of MSPC, most of the industrial plants are still not able to adopt MSPC and really enjoy the benefit of improving product quality. The motivation of this research is to facilitate MSPC implementation in industrial plants while incorporating opinions and needs from the industrial field. MSPC is theoretically proven as a precise statistical process control technique. It can help an industrial plant to monitor the production more properly, detect the abnormal process event more effectively and thus reduce the production cost of a company with a lower defect product rate. However, due to the barriers of implementing MSPC mentioned above, it seems the theoretical knowledge is not successfully transferred into practical field. 1.2. Research Questions and the Objective After knowing the background of the research project, the main research question is defined as follows. “What are the difficulties of multivariate statistical process control (MSPC) implementation and how quality management staff can be supported to facilitate MSPC in a process industry plant?” The main research question is elaborated into five sub-questions for better understanding. 1). What is the essence of MSPC? 2). What are the expectations from quality management staff in the process industry plant? 3). What tools can be provided to make the interpretation of MSPC results easier? 4). What advices can be provided to cope with the time axis problem for MSPC? 5). What recommendations can be provided to quality management staff in the process industry plants? Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 2 Chapter 1. Introduction All the sub-questions will be answered when the research work is complete. The first sub-question basically addresses the background knowledge about MSPC and how MSPC works in the process control. After knowing MSPC, we will investigate the opinions from practice, and the intention is to accommodate the whole research work more workable in practice. We will develop several approaches based on the opinions from practice and these approaches will answer question three and four. For the last question, we will provide general recommendations to quality staff based on the findings from both technical aspect and practical aspect. The research objective of this thesis research is to make recommendations for implementing multivariate statistical process control (MSPC) in a process-industry plant by providing clear interpretations of MSPC and suggestions to quality management staff in the plant. The main contribution of this research work will contain several targets. First, an analysis of Shewhart control charts and various MSPC methods will be concluded. Second, in order to obtain the opinions and needs from the practical field, several interviews with industrial plant staff, statistical process consultants and statisticians will be conducted. Very often the academic theory development does not receive positive feedback derived from the practical implementation because the voice of practical field is missing. Third, by knowing the characteristics of different SPC techniques and combining with the expectations from the practical field, MSPC Implementation Guideline will be developed. This guideline is meant to support the industrial plants to implement MSPC technique. In the guideline, there are two practical tools, which are Method model for (M)SPC and MSPC Diagnosis. The first one supports the practitioners to choose the proper SPC technique under different circumstances. MSPC Diagnosis will serve as a generic Out-of-Control-Action-Plan (OCAP) of the plant while using MSPC. The MSPC Diagnosis can help the plant staffs correctly react when they encounter the out-of-control measurement, and lower the defect product rate. During this research period, these two approaches will be accomplished at the conceptual level. Nevertheless, it is foreseeable that developing profound software which can cover the complex statistic calculation will be a key to lower the barrier for implementing MSPC in a plant. The approaches proposed in this research can serve as functional specifications of software development. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 3 Chapter 1. Introduction Research Background SPC Research Objective Shewhart Control Chart Gap between Academic and Practical Fields Research Questions MSPC Control Chart Industrial Expectations Interviews MSPC Plan Case Briefing MSPC Training MSPC Implementation Team Approach Result Management Involvement Process Performance Improvement Conclusions Recommendations Future Research Figure 1-1 Structure of the thesis. The structure of this thesis is shown in Figure 1-1 and details are described as follows. Chapter 2 will introduce some statistical process control (SPC) techniques. Shewhart control charts will be briefly introduced and multivariate statistical process control (MSPC) will be emphasized with several approaches. Hotelling’s T2 statistic and several successive approaches such as T2A and SPE plot, MYT T2 decomposition and Principal Component Analysis (PCA) application will be addressed and discussed. Chapter 3 will involve the views of practice. The opinions and the expectations from industrial plant staff, statistical process consultants, and statisticians will be concluded and further incorporated into our research output. MSPC Implementation Guideline, including the scheme of Method model for (M)SPC and MSPC Diagnosis will be illustrated in Chapter 4 with detailed explanation. In Chapter 5, the effectiveness of these two approaches will be validated with a case application. At the end of this thesis, conclusions and recommendations will be presented in Chapter 6. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 4 Chapter 2. Statistical Process Control Chapter 2. Statistical Process Control The aim of statistical process control (SPC) is to achieve higher quality of final product and lower the production loss due to defect product. Process monitoring with control chart is a basic tool of statistical process control. It monitors the behavior of a production process and signals the operator to take necessary action when abnormal event occurs. A stable production process is the key element of quality improvement. In this chapter, the traditional control chart – Shewhart control charts, which is a univariate statistical process control technique will be introduced. After that, a multivariate statistical process control (MSPC) technique – Hotelling’s T2 method and its advantages/drawbacks will be discussed. With the same idea of Hotelling’s T2 method, an adjusted approach T2A and SPE plot will be introduced as well. Knowing the problem of interpretation of the result of Hotelling’s T2 method, two diagnostic methods: (1). Application of Principal Component Analysis (PCA) and (2). MYT T2 decomposition will be reviewed with an example. In the end of chapter, comparison and discussion will be made for these methods. 2.1. Shewhart Control Charts The originator of statistical process control chart is Dr. Walter A. Shewhart. The basic idea of Shewhart control chart requires an analyst to take samples from the process periodically and calculate a statistic to summarize the process behavior. The measurements are plotted on the chart against time or observation series and compared to control limits drawn on the chart (Stephen, et al. 1999). A generic Shewhart control chart is shown in Figure 2-1. The center line represents the expected value of the quality characteristic during in-control process. The upper control limit (UCL) and lower control limit (LCL) are chosen based on the nature the process behavior, which means only a certain probability that the process falls within in-control limits (please see 2.1.1 Control limits). The control limits are directly calculated from the process data. It should be noted that control limits are not specification limits defined by customer. Therefore, in-control process does not mean that the product meets the specification limits, it only means that the process behavior is consistent and predictable. From the interview with industrial statisticians, we discovered that it is sometimes a common mistake that the quality people take the customer-defined specification limits as control limits while applying Shewhart control charts. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 5 Chapter 2. Statistical Process Control Figure 2-1 A generic Shewhart control chart. One of Dr. Walter A. Shewhart’s (Shewhart, 1931) fundamental concepts was that, the variation of a process results from two sources. One is called common cause, which is inherent in the production system and it is not possible to remove the common cause from the process unless some changes of the existing process system have been taken. The variation from common cause is a very minor fluctuation and does not harm the final product quality. The other is called special (assignable) cause, which is resulted from several particular reasons (e.g. problem with raw material, operator mistakes, machine failures, etc.) and special cause may lead to serious damage to the final product quality and cause the loss of a company. The Shewhart control charts serve as a tool to detect the abnormal event caused by special causes and signal the operator to analyze the problem. 2.1.1. Control Limits A point falling within the control limits means it fails to reject the null hypothesis that the process is statistically in-control, and a point falling outside the control limits means it rejects the null hypothesis that the process is statistically in-control. Therefore, the statistical Type I error α (Rejecting the null hypothesis H0 when it is true) applied in Shewhart control chart means the process is concluded as out-of-control when it is truly in-control. Same analog, the statistical Type II error β (Failing to reject the null hypothesis when it is false) means the process is concluded as in-control when it is truly false. According to Engineering Statistics Handbook (NIST/SEMATECH e-Handbook of Statistical Methods, http://www.itl.nist.gov/div898/handbook/, retrieved on 01/may/2005), the UCL is pictured when the probability of a measurement falls above of UCL is 0.001. The same situation holds for LCL. Therefore, the probability will be 0.002 when a measurement falls either outside UCL or LCL, and probability of 0.002 is practically considered acceptable quality control. Compared with the normalized standard distribution probability, the probability of a measurement falling outside of the limit which locates 3 sigma (standard deviation) away from average is 0.00135. For both sides, the probability is 0.0027 of a measurement falling Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 6 Chapter 2. Statistical Process Control outside of UCL/LCL. Therefore 3 sigma has become a customary distance between central line and UCL/LCL and generally it gives good results in practice. Average Run Length (ARL) The performance of control charts can also be characterized by their average run length. Average run length is the average number of points that must be plotted before a point indicates an out-of-control condition (Montgomery, 1985). We can calculate the average run length for any Shewhart control chart according to, ARL = 1 p where p is the probability that an out-of-control event occurs. Therefore, a control chart with 3 sigma control limits, the average run length will be ARL = 1 1 = = 370 p 0.0027 This means that if the process remains in-control, in average, there will be one false alarm every 370 samples. 2.1.2. Patterns of Process Behavior Apart from all the measurement should fall with the control limits, the process can be viewed as in-control when there is no systematic pattern shown in the process behavior. Systematic patterns occurring in Shewhart control charts have often been interpreted as indicators of extraneous sources of process variation (Mason, et al. 2003). The process will be improved if the causes of systematic pattern in the process are diagnosed and further eliminated. Typical patterns are shown in Figure 2-2. Cyclic pattern may be caused by systematic environment change, such as seasonal temperature or operators shifting. Trend pattern is usually due to wearing out of a tool/machine or catalyst deterioration. A shift in process level may be caused by the feeding of new material, or by the operation run by a new worker. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 7 Chapter 2. Statistical Process Control Figure 2-2 Typical systematic patterns. The Western Electric Handbook (1956) provides a set of guidelines to detect the systematic patterns in the process. A brief summary is shown below. A process is considered as out-of-control if any of the following conditions holds: 1). One point falls outside the 3-sigma control limits (beyond Zone A). 2). At least two out of three consecutive points fall on the same side of the center line, and are beyond the 2-sigma control limits (in Zone A or beyond). 3). At lease four out of five consecutive points fall on the same side of the center line and are beyond the 1-sigma limits (in Zone B or beyond). 4). At least eight successive points fall on the same side of the center line. Figure 2-3 The Western Electric run rules. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 8 Chapter 2. Statistical Process Control 2.1.3. Control Charts for Attributes We often come to the situation when we are not able to measure the quality characteristic of a product, for example, checking the surface scratches of a product or proportion of malfunctioning lamps. The result of checking can be classified into whether conforming or nonconforming to the specification on the quality characteristic. In this case, control charts for attributes should be applied. The characteristics of each control chart and its application will be introduced as follows. The control limits are constructed according to customary 3-sigma distance away from the center line. Comprehensive formulas for computing the centerline and control limits of control charts of attributes can be found in Appendix A. p-chart The p-chart graphs the proportions of defective items from successive subgroups. It tells us the defect rate of the product. It should be noted that the sample size should be large enough to contain defective products; otherwise the control chart will lose the meaning of detection if most of the p values from the samples are zero. np-chart The np-chat is slightly different from p-chat. Instead of plotting the proportions of defective items, the number of defectives np is plotted. In order to make the number of defectives comparable, it is important that the sizes of sample have to be the same. For shop floor operators, the information from np-chart is more straightforward than p-chart and easier to understand. C-chart The c-chart is applicable when the large product is inspected. The quality can be monitored in terms of counting the number of nonconformities on each product (sample size is one). U-chart The u-chart is a modification of the c-chart. The number of nonconformities per unit (ui = ci / ni) is plotted, so the sample size does not need to be one. The probability of the occurrence of nonconformity can be increased with larger sample size n to avoid too many detecting results (u value) are zero. 2.1.4. Control Charts for Variables A single measurable quality characteristic, such as a dimension, weight, or volume, is called a variable. Control charts for variables are used extensively. They usually lead to more efficient control procedures and provide more information about process performance than attributes control charts (Montgomery, 1985). When a variable is monitored, it is a standard practice to control both the mean and the variability of the variable. The mean of variable is monitored with Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 9 Chapter 2. Statistical Process Control x-bar chart (mean chart) and the variability of the variable is monitored with S-chart (standard deviation chart) or R-chart (range chart). The x-bar chart can tell us whether the process is stable with respect to its level. S-cart and R-chart can tell us whether the variability of the process is stable over time. Significant shifting of the mean and the unusual large variability are the indications of special causes, which need to be detected and eliminated. Therefore, it is important to monitor both simultaneously. The control limits are constructed according to customary 3-sigma distance away from center line. Traditionally, quality-control engineers have preferred the R-chart to the S-chart because of the simplicity of calculating R from each sample. However, the R-chart is relatively insensitive to small or moderate shifts for small sample size. Thus, in the situation that tight control of process variability is needed, moderately large sample sizes will be required, and the S-chart should be used (Montgomery, 1985). I-chart In the situation that the production rate is very slow, and difficult to accumulate more than one sample unit before analysis, I-chart (Individual chart) should be applied. Again, the mean and the variability of the process should be monitored simultaneously. The mean of a process is computed directly from the mean of the entire individual sample and the variability is computed from the moving range. Due to the sample size is one, the process variability is estimated with the moving range MR=│Xi – Xi-1│, which is the absolute value of the difference between two adjacent observations. Comprehensive formulas for computing the centerline and control limits of control chart of variables can be found in Appendix A. CUSUM Control Chart and EWMA Control Chart Apart from various types of Shewhart control chart, there are two effective alternatives, which should be shortly introduced. The first one is cumulative-sum (CUSUM) control chart (Page 1954) and the other is exponentially weighted moving-range (EWMA) control chart (Roberts 1959). CUSUM control chart accumulates the deviations of each measurement from the center line, while the weighted average gives more weight to the more recent measurements and less to those in the past. Since CUSUM contains all the information from the whole sequence of measurements not only the last one, it is more effective than Shewhart control charts on detecting small process shifts. Similar idea for exponentially weighted moving-range (EWMA) control chart. The only difference is that for EWMA, the weights that are given to the measurements decrease geometrically with the age of the sample mean. Additional reference can be referred to D.C. Montgomery (Introduction to Statistical Quality Control). 2.2. Multivariate Statistical Process Control The Shewhart control charts have been widely applied in a variety of industries because it is very easy to implement and the information generated from the Shewhart control charts is also easy Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 10 Chapter 2. Statistical Process Control for plant staff to understand. However, monitoring each process variable with separate Shewhart control chart ignores the correlation between variables and does not fully reflect the real process situation. Nowadays, the process industry has become more complex than it was in the past and inevitably that number of process variables need to be monitored has increased dramatically. Very often, these variables are multivariate in nature and using Shewhart control charts becomes insufficient. Figure 2-4 demonstrates how misleading information could be generated from Shewhart control chart in the multivariate circumstance. Figure 2-4 An example of misleading information generated from Shewhart control chart. Assuming a doll production, and there are two quality variables with positive correlation between them (for the convenience of visual illustration, product quality variable Height & Weight are used as of process variables). The multivariate control chart on the right side is simply two Shewhart control charts superimposed together according to vertical and horizontal axes. Due to the positive correlation between H and W, it is expected that the measurement plots will locate within the elliptical region. Doll 2 and 3 are considered as in-control process because they fulfill the control limits of Shewhart control chart as well as the multivariate control chart (the elliptical region). While doll 4 will be easily detected as an out-of-control event since it falls outside the UCL of Shewhart control chart. The critical doll is plot 3, which falls outside of the expected region (the ellipse) while both Shewhart control charts appear to be in-control. In multivariate circumstance, an out-of-control signal can be caused by (1). Extraordinary value of a variable or a set of variables, (2). Due to the relationship between two or more variables which contradicts the pattern established by the historical data or (3). A combination of the former two causes. Very often the process variables are not independent. They sometimes influence each other following certain predictable patterns. For example, one variable should become larger while the other variable becomes larger (positive relation), whereas one variable should become smaller while the other variable becomes larger (negative relation). By using Shewhart control charts is not able to signal the process is out-of-control when the relation between process Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 11 Chapter 2. Statistical Process Control variables deviates from its predicted pattern. In this case, the plant staff may lose the chance to detect problematic process and further investigate the problem in time. 2.2.1. Hotelling’s T2 Statistic Hotelling H. (1931) can be viewed as the originator of multivariate control charts. Hotelling proposed a concept of generalized distance between a new observation to its sample mean. We first illustrate how this method works with a bivariate case. Assuming these x1 and x2 are distributed according the bivariate normal distribution. Referring to Figure 2-5, say X1 and X2 are the mean, σ1 and σ2 are the standard deviation of these two variables respectively. The covariance σ12 is used to estimate the dependency between x1 and x2. The generalized distance between point A and its mean can be calculated as: X0 2 = 1 ⎡ s11 ( x 2 - x 2 )2 − 2( x 2 - x 2 )( x1 - x1 ) + s22 ( x1 - x1 )2 ⎤ ⎦ s11 s22 − s212 ⎣ This statistic follows the Chi-square distribution with two degrees of freedom. An ellipse can be graphed with the x1 and x2 in this equation. Moreover, all the points lying on the ellipse will generate the same Chi-square statistic. As a consequence, every observation can be determined whether its generalized distance exceeds the ellipse by comparing X02 and X22,α ,where X22,α is the upper α percentage point of the Chi-square distribution with 2 degrees of freedom. The observation will be considered as out-of-control if X02 > X22,α. Figure 2-5 A generic bivariate Hotelling’s T2 control region. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 12 Chapter 2. Statistical Process Control With the same concept of the generalized distance, it can be extended from bivariate to a ' multiple p variables. Let Xi = (Xi1, Xi2 ,...., Xip ) represent a p dimensional vector of measurements made on a process time period i. The value Xij represents an observation on the jth characteristic. Assuming that when the process is in control, the Xi are independent and follow a multivariate normal distribution with mean vector µ and covariance matrix Σ. Normally µ and Σ are unknown, but we can use X and S estimated from a historical data set with n observations. Phase I and Phase II The application of Hotelling’s T2 statistic shall be categorized into two phases. Phase I tests whether the preliminary process was in control and phase II tests whether the future observation remains in-control (Alt, 1985). Phase I operation refers to the construction of in-control data set. Same idea as Shewhart control chart, control limits are estimated from a period of in-control data. To obtain this in-control data, the raw data set needs to be purged. For instance, the outliers need to be removed and the missing data needs to be substituted with an estimate. During phase I operation, Hotelling’s T2 statistic is calculated for each measurement and compared to the control limit, which will follows Chi-square distribution (according to Richard, A.J. & Dean, W.W., 2002.) T 2 = (Xi - X)' S-1(Xi - X) ~ X2α,p (Chi - square distribution) (eq. 2 − 1) Also other research shows that the control limit follows Beta distribution (Mason, Young & Tracy, 1992). T 2 = (Xi - X)' S-1(Xi - X) ~ (n-1)2 B p n −p −1 ( α, , ) n 2 2 (eq. 2 − 2) n: number of preliminary observations Both control limits will be approximate when the number of observations is large. The control limit based on Chi-square distribution is established on the assumption that X and S are true values µ and Σ, which is just an approximate situation (Mason, Young & Tracy, 1992). Beta distribution is more precise and is a recommendable choice. After purging the raw data with Hotelling’s T2 statistic, the in-control data set is ready for monitoring future observations which is termed as phase II operation. The control limit for determining future observation is different from the one in phase I. It follows an F distribution with p and (n-p) degrees of freedom. p(n+1)(n-1) F(p,n −p, α ) n(n − p) n: number of preliminary observations T 2 = (Xi − X)' S−1(Xi − X) ~ Master Thesis: Multivariate Statistical Process Control in Industrial Plants. (eq. 2 − 3) Page 13 Chapter 2. Statistical Process Control Where sample mean is X = (X1, X2 ,...., Xp )' and the covariance of sample ⎡ s11 s12 s13 .... s1p ⎢ s22 s23 .... s2p ⎢ S=⎢ O ⎢ ⎢⎣ spp ⎤ ⎥ ⎥ ⎥ ⎥ ⎥⎦ The idea of using Hotelling’s T2 statistic in phase I and phase II is the same. Each measurement is examined whether it is out-of-control by checking if it deviates extraordinarily from its sample mean. It should be reminded to choose the correct upper control limit on different purposes. The Hotelling’s T2 statistic can be extended for more than two variables. Instead of a 2-dimensional ellipse control region, the result will be presented in a similar way as Shewhart control chart. The T2 statistics calculated from all the observation will be plotted in a chart against time or observation serious and compared to the upper control limit. Figure 2-6 is a generic T2 control chart. It should be noticed that there is no center line and the lower control limit is set to zero, because the meaning of T2 statistic is a generalized distance between the observation and its sample mean. Figure 2-6 A generic T2 control chart. Case Demonstration We will demonstrate how Hotelling’s T2 statistic helps us to determine whether a measurement is in-control with an example. The data set was used in Hawkins’ paper (1991). Data set can be found in Appendix C. The data contained 50 measurements of 5 variables and measurements 1 to 35 were considered as in-control process. In this case, the data set was already purged, so we only perform the phase II operation-monitoring the future observations. An upward shift of 25% of a standard deviation was introduced to X5 while the marginal standard deviation of X1 was introduced by 50 % for measurement 36 to 50. X and S are estimated from the measurement 1 to measurement 35. As we know Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 14 Chapter 2. Statistical Process Control that the Hotellings’s T2 statistic will follow an F distribution with p and (n-p) degrees of freedom, the upper control limit is calculated as, p(n+1)(n-1) F(p,n −p, 0.05 ) n(n − p) 5(35+1)(35-1) = * 2.53 35(35 − 5) = 14.75 UCL = Recall the idea from Figure 2-5, the value of UCL means the generalized distance between the mean (the center point) of the measurements obtaining from a period of in-control process to the control limit (the ellipse). So now we can compute Hotelling’s T2 statistic (eq. 2-3) for measurement 36 to measurement 50 to determine its status. A measurement will be considered as out-of-control if its Hotelling’s T2 statistic is larger than UCL. Figure 2-7 shows the plot for each measurement, and measurement 48 is detected as an out-of-control situation. Figure 2-7 T2 control chart of measurements 36 to 50. 2.2.2. T2A & SPE Plot The Hotelling’s T2 statistic is very effective and easy to understand the result. However, using T2 with highly correlated the variables; the covariance matrix is often very ill-conditioned. When the number of variables is large, the covariance matrix is often nearly singular and may not be inverted (Kourti & MacGregor, 1995). Without covariance inversion matrix, the Hotelling’s T2 statistic is not possible to obtain. With this concern, another approach was proposed by Kourti and MacGregor (1995). The details will be explained as follows. The traditional T2 equation Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 15 Chapter 2. Statistical Process Control T 2 = (Xi − X)' S−1(Xi − X) can be expressed (Mardia, Kent and Bibby, 1989; Kourti and MacGregor, 1994) as, q T2 = ∑ i =1 q q A t i2 t2 t2 t2 = ∑ i2 = ∑ i2 + ∑ i2 λi i =1 si i =1 si i = A +1 si ti = Pi (X − X) where λi and Pi are the eigenvalues and eigenvectors of principal components that generated from the original data. So the Hotelling’s T2 statistic based on the first A principal components is, A TA 2 = ∑ i =1 t i2 si2 When TA2 does not utilize all the principal components, it is just an approximated value of T2; therefore it will be equivalent to T2 if all the principal components are utilized. To construct TA2, it is not necessary to obtain the inversion of covariance matrix anymore; besides, the dimensionality can be reduced as well. Due to TA2 is only an approximated value of T2, it only can detect whether there is an abnormal variance occurs in the plane constructed with A principal components. If a totally new type of special even occurs which was not present in the reference data used to develop the in-control PCA model, then new principal component will appear and the new observation will move off the plane (Kourti & MacGregor, 1995). So we need another support which is squared prediction error (SPE). The SPE is the squared perpendicular distance of an observation xi from the projection space and it tells us how close the observation xi is to the space constructed with A principal components. To determine the status of a new observation, both TA2 and SPE are needed. A t 2new,a a =1 sa 2 TA 2 = ∑ UCL = A ⎡ ' P (X − X) ⎤ = ∑ ⎢ a new ⎥ sa a =1 ⎣ ⎢ ⎦⎥ 2 (eq. 2 − 4) p(n+1)(n-1) A(m+1)(m-1) F(p,n −p, α ) = F( A,m −Α, α ) n(n − p) m(m − A) 2 n SPE x,new = ∑ (Xnew,j − Xnew,j ) (eq. 2 − 5) j =1 Xnew = a=A ∑ (t a =1 P) new,j a Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 16 Chapter 2. Statistical Process Control Case Demonstration We will apply the same data set that we used in the previous sections for demonstration. Since the eigenvalues of the first two principal components are 87% of the total eigenvalues, which means the first two principal components already explained most of the variance from the original data. Under such circumstance, the TA2 of this case will be computed with two principal components. Applying the formula above, TA2 and SPE are computed (eq. 2-4 & eq.2-5) and plotted for measurement 36 to measurement 50 in Figure 2-8 and Figure-2-9. Figure 2-8 TA2 control chart based on first two principal components. Figure 2-9 SPE chart based on first two principal components. From the T2 chart, the measurement 48 appears to be out-of-control. SPE chart also shows measurement 48 has relative higher value than others, which means this measurement is far from the projection model constructed from in-control historical data. The result from this approach is similar with the one applied Hotelling’s T2 approach, but it avoids the problem of inversion of covariance matrix. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 17 Chapter 2. Statistical Process Control Both Hotelling’s T2 statistic and T2A & SPE plot offer a great support on the deficiency of Shewhart control chart, which cannot monitor the correlation between variables. In addition, both approaches can reduce a large amount of individual Shewhart control chart into a synthetic control chart. However, since the T2 statistic (either using Hotelling’s T2 statistic or T2A & SPE plot) is the synthetic statistic value generated from more than one process variables, it is difficult for user to determine which variable or set of variables is responsible when the abnormal event occurs. Without knowing this information, it is difficult for plant staff to search the root causes of the abnormal event and further eliminate them. In this case, Hotelling’s T2 statistic and T2A & SPE plot approaches are able to signal the operator when something goes wrong in the process, yet they are not able to tell the operator what is wrong or how exactly the process goes wrong. In the following section, two approaches will be discussed for diagnosing responsible variable(s) once abnormal event is detected. 2.3. Diagnostic Approaches for Hotelling’s T2 Method In order to support Hotelling’s T2 statistic and T2A & SPE plot approaches to identify the source of an abnormal signal, two approaches have been proposed. First, Mason, Young & Tracy proposed a decomposition approach to breakdown T2 into orthogonal components (1995, 1997). Second, Kourti and MacGregor (1995, 1996) provided another diagnostic approach based on Principal Component Analysis to identify the responsible variable(s) for abnormal measurement. These two approaches will be further explained with a case demonstration. 2.3.1. MYT T2 Decomposition Mason, Young and Tracy (1995, 1997, 1999) proposed an approach (hereafter is referred as MYT approach) to decompose the Hotelling’s T2 statistic into orthogonal components. Findings from Hotelling’s T2 can thus be interpreted in a way that most people can follow. The MYT approach is applied after the abnormal measurement is detected by either Hotelling’s T2 statistic or T2A & SPE plot for identifying the responsible variable or set of variables. For a p-dimensional vector, one form of the MYT approach can be expressed as, T 2 = T12 + (Τ2 2•1 + .... + T 2p•1,2,....,p −1 ) 2 (x j − x j•1,2,....,j−1 ) (x − x1 ) = 1 2 + s1 s2 j•1,2,....,j−1 2 j = 1, 2, ... , p The first term T12 is an unconditional Hotellings’ T2 for the first variable of the measurement. The rest of the terms are referred as conditional terms. It should be noted that the ordering of the Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 18 Chapter 2. Statistical Process Control individual conditional terms is not unique. There are p! different partitionings that can generate the same overall T2 statistic (Mason, Young & Tracy, 1995). For example, we can start with selecting any one of the p variables. Then we can choose any of the (p – 1) remaining variables to condition on the first selected variable. Next we can choose any of the remaining (p – 2) variables to condition on the first two selected variables. Iterating the same procedure will generate all the decomposition equations which compose the same over T2 statistic. Taking a case of three variables as an example, it can be decomposed as, T 2 =T12 + Τ 2 2•1 + Τ 23•1,2 =T12 + Τ 23•1 + Τ 22•1,3 =T2 2 + Τ 23• 2 + Τ 21•2,3 =T2 2 + Τ 21•2 + Τ 23•1,2 =T3 2 + Τ 21•3 + Τ 22•1,3 =T3 2 + Τ 22•3 + Τ 21•2,3 It is obvious that with the increase of the number of variables, the number of terms will also increase dramatically which makes the computation become troublesome. Nevertheless, the two terms of greatest interest are often the unconditional term and the term containing the adjusted contribution of one of the variables after adjusting for the other (p – 1) variables (Mason, Young and Tracy, 1995). Unconditional Term The unconditional term has a similar function of a univariate Shewhart control chat. It calculates the squared standardized variance of jth variable. A signal will occur if jth variable is too far away from the sample mean. T2j will follow an F distribution which can be used as upper control limit. Tj = 2 (x j − x j ) sj 2 2 ∼( n +1 ) F(1,n-1,α ) n (eq. 2 − 6) Conditional Term The conditional term is a standardized observation of the jth variable adjusted by estimates of the mean and variance from the conditional distribution associated with xj•1,2,…j-1 . The most important function of conditional term is that it measures whether the jth variable is consistent to the relationship pattern with other variables established from historical in-control Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 19 Chapter 2. Statistical Process Control data. The conditional term T2j˙1,2,…j-1 will follow an F distribution which can be used as upper control limit. T 2 j•1,2,....,j −1 = (x j − x j•1,2,....,j−1 ) s 2 j•1,2,....,j −1 2 ⎛ (n+1)(n-1) ⎞ ∼⎜ ⎟ F(1,n-k-1,α) ⎝ n(n-k-1) ⎠ κ=(j-1) (eq. 2-7) T2j˙1,2,…j-1 can be re-expressed as (Mason, Young and Tracy, 1997), T 2 j•1,2,....,j−1 = (x j − x j•1,2,....,j−1 )2 s2 j (1 − R 2 j•1,2,....,j−1 ) (eq. 2-8) The numerator is the squared residual between the observation and the predicted point based regressed by the variables x1, x2, … xj-1. R2j•1,2,…j-1 is the squared multiple correlation coefficient between xi and x1, x2, … xj-1. From the equation eq.2-8, it is noticed that the conditional T2 term will become large if xj is significantly far from what is predicted from the historical data, unless the Rj•1,2,…j-1 is close to 1. Reduced Computation Scheme As stated previously the number of the unique decomposition terms will increase dramatically when the number of variable increases. Table 2-1 provides more detailed information. Table 2-1 Unique MYT decomposition terms. (cited from Mason, Young & Tracy, 1997) Therefore, a reduced computation scheme was proposed also by Mason, Young & Tracy (1997). Here we summarized this reduced computation scheme as follows. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 20 Chapter 2. Statistical Process Control Step 1. Compute the individual statistic Ti2 (according to eq. 2-6) for every component of the X vector. The variables with significant T2 statistic are out of individual control and it is not necessary to check how they relate to other variables. Check whether the subvector with remaining k variables produces a signal. Sept 2. (Optional but useful for very large p). Examine the correlation structure of the subvector. The variable with very weak correlation (0.3 or less) can be removed. Step 3. If the subvector still produces a signal, then compute all the T2i,j terms (according to eq. 2-8). T2i,j terms tell us something is wrong with bivariate relationship, if T2i,j is significant. Continue to check the T2 statistic for the remaining subvector. If no signal occurs, then it is concluded that the individual variable from step 1 and the relationship between the bivariate are the sources of the abnormal measurement. Step 4. If the subvector of remaining variables still produces a signal, then compute all the T2i,j,k terms. Follow the same rule from previous steps and examine all the conditional terms. Step 5. Repeat the same procedures until the T2 statistic of the remaining subvector is not significant. Case Demonstration We will apply the same data set that we used in the previous sections for demonstration. The abnormal situation is detected with Hotelling’s T2 statistic approach. 1). Hotelling’s T2 statistic was computed with eq. 2-3 and plotted for measurements 36 to measurement 50. Measurement 48 is detected as an abnormal measurement. Please refer to Figure 2-7 T2 control chart of measurement 36 to 50. 2). Table 2-2 shows all the individual Ti2 calculated from eq. 2-6 and compared with upper control limit. . n +1 ) F(1,n-1,0.05) n 35 + 1 =( ) F(1,35-1,0.05) 35 = 4.248 UCL = ( Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 21 Chapter 2. Statistical Process Control Table 2-2 Individual Ti2 and its status. 3). Remove x1and check whether the subvector signals or not. The result shows the subvector is still significant. T 2 - T12 = 22.92 - 7.61 = 15.31 > 14.75 4). It is concluded that not only x1 is problematic; the relationship between variables in the subvector is also the possible cause. So we continue to check bivariate statistic T2i,j. with eq. 2-8. A summary table can be found in Table 2-3. ⎛ (n + 1)(n - 1) ⎞ UCL = ⎜ ⎟ F(1,n-k -1,α ) ⎝ n(n - k - 1) ⎠ ⎛ (35 + 1)(35 - 1) ⎞ =⎜ ⎟ F(1,33,0.05) ⎝ 35(35 - 1- 1) ⎠ = 4.39 Table 2-3 Bivariate conditional term and its status. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 22 Chapter 2. Statistical Process Control T25˙4 is found to be significant, which means the relation between x4 and x5 is problematic. So we may conclude that both x4 and x5 are potential causes of the abnormal observation. 5). Remove x4 , x5 and check whether the subvector signals or not. The result shows the subvector is not significant anymore. So the computation can stop. T 2 - T12 - T4 2 - T5 2 = 22.92 - 7.61 - 0.58 - 3.87 = 10.86 < 14.75 So far, we can conclude that x1 is individually out-of-control. Moreover the meaning of significant value of T25˙4 is that the x5 (conditioned by x4) deviates from the variable relation pattern established from historical data. Finally, x1 and x5 are determined as the responsible sources of abnormal measurement 48, which is within our expectation. 2.3.2. T2 Diagnosis with Principal Component Analysis (PCA) Another diagnostic approach is based on the idea of a well-known statistic technique-Principal Component Analysis (PCA). The most noticeable advantage of Principal Component Analysis (PCA) is that it can generate a new set of orthogonal principal components (principal component is a linear combination of all original variables) based on the original data set and the information from the original data can be explained by fewer components. The complete set of principal components can reproduce the total variance of the original data set. However, most of the variance can be captured by the small number k principal components and thus the dimensionality can be reduced. Once a PCA model is constructed based on an in-control historical data, the original variables are considered simultaneously and the relation between variables are also captured. A new observation can be detected as out-of-control if it significantly deviates from the PCA model. The key elements of principal components are the eigenvectors (pi) and eigenvalues (λi) generated from the covariance matrix S of in-control historical data. Eigenvectors (pi) serve as the axes of principal components, while eigenvalues (λi) are the variances of the principal components. The T2 statistic can be expressed in terms of the principal components (Mardia, Kent, and Bibby 1979, Jackson 1991) n ta2 t2 = ∑ a2 a =1 λ a a =1 sa n T2 = ∑ n t a = Pa' (x - x) = ∑ pa,j (x j - x j ) (eq. 2 − 9) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 23 j =1 Chapter 2. Statistical Process Control where ta are the scores from the principal component transformation, and xj is the jth quality characteristic. In addition, the normalized PCA scores (ta/sa) can be calculated from each principal component and we can compare which principal component(s) contributed most to the abnormal measurement once it is detected by Hotelling’s T2 or T2A & SPE plot approaches (see section 2.2.). The principal component(s) with higher normalized scores (ta/sa) can be further investigated with contribution plotting (MacGregor, et al. 1994). The contribution of each variable to a particular principal component is, pa,j (x j − x j ) (eq. 2 − 10) If we plot the contribution of every variable, those with higher values are more likely to be responsible for the abnormal measurement and need to be investigated. Case Demonstration We will apply the same data set that we used in the previous sections for demonstration. The abnormal situation is detected with Hotelling’s T2 statistic approach. 1). Hotelling’s T2 statistic was computed with eq. 2-3, and plotted for measurements 36 to measurement 50. Measurement 48 is detected as an abnormal measurement. Please refer to Figure 2-7 T2 control chart of measurement 36 to 50. 2). Normalized PCA scores (ta/sa) are calculated with eq. 2-9, and plotted in Figure 2-10. The normalized score of principal component 2 and 3 are relatively high. In addition, the limits for the normalized scores are roughly used as a guide, and 2.7σ would be equivalent to Bonferroni-type limits (I.e., replace α/2 with α/2n) 95% confidence on type I error for five variables. Figure 2-10 Normalized PCA scores. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 24 Chapter 2. Statistical Process Control 3). Knowing normalized score 2 and 3 contributed most to this abnormal measurement, we can construct the contribution plot (using eq. 2-10) of 5 variables for score 2 and score 3 (see Figure 2-11 and 2-12) to see which variable or set of variables contributed most. Figure 2-11 Variable contribution plot of principal component 2. Figure 2-12 Variable contribution plot of principal component 3. From the contribution plot of score 2, it is clear that x1 contributed most among all variables. In contribution plot of score3, x5 contributed most and x1 also had relatively high value. It should be noted that score 3 was negative in the score plot. Thus, in contribution plot of score 3, we should only look at the variables with negative value because the positive contributions only make the score smaller. Therefore, it is suggested that variables with high contributions but with the same sign as the score should be investigated (Kourti & MacGregor, 1996). So far, we can conclude that x1 and x5 are the variables that need further investigation for this abnormal measurement. It can also happen that more than one score with high value, for instance, there are two scores with relatively high value in our case demonstration. Overall average contribution per variable is suggested (Kourti & MacGregor, 1996). The steps of constructing overall average contribution are summarized as below. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 25 Chapter 2. Statistical Process Control 1). Select the k normalized scores with high values. In this case, score 2 and score 3 were selected. 2). Calculate contribution of a variable xj in the normalized score. Cont a,j = ta pa,j (X j − X j ) λa Conta, j is set to zero if it is negative. (i.e., the sign of variable contribution is opposite to the value of score) 3). Calculate the total contribution of variable xj. k CONTj = ∑ (cont a,j ) j =1 The overall average contribution per variable generates an overview contribution of each variable in one plot, which is very convenient. Here again, we see that variable x1 and x5 are responsible for the abnormal situation. Figure 2-13 Overall average contribution per variable. 2.4. Approaches Discussion In order to clearly demonstrate a structure of multivariate statistical process control approaches included in this research, a tree diagram is provided in Figure 2-14. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 26 Chapter 2. Statistical Process Control Legend : Decision point MSPC approaches Decision option Action Stage I Yes Inversion of covariance matrix available? T2 control chart No TA2 and SPE control chart Stage II PCA diagnosis and overall contribution plot. MYT T2 decomposition PCA diagnosis and overall contribution plot. MYT T2 decomposition Figure 2-14 A structure of multivariate statistical process control approaches. After the discussion of the mechanisms of several approaches (Hotelling’s T2, T2A & SPE plot, MYT T2 decomposition and PCA application,) presented above, a general discussion on their applications will be made. With better understanding of their characteristics, we can make better choice among them under different circumstances. The discussion will be categorized into two stages: Stage I focusing on the detection the abnormal measurement, and stage II focusing on the diagnosing the sources of abnormal measurement. Stage I – Detecting the Out-of-Control Event 1). As we have discussed in Chapter 2.2., MSPC can not only monitor the status of variables also monitor the relationship between variables, particularly when variables are highly correlated of course. Whereas Shewhart control charts are not possible to monitor the relationship between variables. 2). Hotelling’s T2 statistic is an effective approach for detecting abnormal measurement in the process. Yet, when the number of variables is large, the covariance matrix is often nearly singular Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 27 Chapter 2. Statistical Process Control and may not be inverted (Kourti & MacGregor, 1995). T2A and SPE plot approach can cope with the problem of covariance matrix inversion and thus would be an alternative choice to Hotelling’s T2 method. 3). Large process systems are often comprised with several process units in it. Breaking down the process system into logical sections, which have highly correlated variables within section but less correlation between sections can be a recommendable idea. Analyzing smaller sections with Hotellin’g T2 statistic separately can reduce the complexity of a large number of variables; in addition, it would become much easier when diagnosing the sources of abnormal measurement. Stage II – Diagnosing the Sources of an Out-of-Control Event 4). MYT T2 decomposition which breaks down the overall Hotelling’s T2 into orthogonal component provides useful information to identify the sources of abnormal measurement. However, even with the reduced computation scheme, the remaining numerous computations, especially when the number of variable is large, may still discourage practitioner to apply it. Programming the computation and breaking down the process system are both considerable ideas to conduct MYT T2 decomposition approach. 5). Diagnosing abnormal measurement with the normalized scores principal components and contribution plot is clear and effective. In addition, the idea of finding high score of principal components and further investigating the variables with high contribution to the selected principal component is quite straightforward. The implementation can be further facilitated with simple calculation sheet such as Excel, which is considered as an easier approach. Another advantage is that the result can be easily understood and communicated between different levels of practitioners with graphical presentation. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 28 Chapter 3. Industrial Practice Chapter 3. Industrial Practice Multivariate statistical process control (MSPC) techniques, including the approaches that we have discussed in Chapter 2, have been discussed and published in academic journals for years. These approaches appear to be effective and may have great opportunity to improve the quality of industry to a higher level. Yet, MSPC has not become a popular technique in industrial plants as expected. It turns out that the link between theory and practice remains missing. Due to this concern, apart from paying attention on theoretic development, we would like to investigate the perspective of MSPC from practical field and why they feel difficult to adopt MSPC. We believe that the barrier of MSPC implementation can be lowered if the opinions of practice are taken into account. In this chapter we will present how the interviews with people from practical fields were conducted. The interviewees are from, for example in manufacturing plants, SPC consultant companies, and in academic institute. Conclusions of the interviews will be made and the information from interviews will be further incorporated in our research developments. 3.1. Selection of the Interviewees In order to obtain objective and direct information regarding the application of MSPC, three different target groups, namely, statistical process control (SPC) consultants, industrial statisticians, and academic statisticians, were selected for the interview. We would like to know from industrial statistician what current statistical process control techniques are being used in the process plants and the perspective on MSPC technique. We also expect that the SPC consultants have more practical knowledge on SPC and MSPC, and we may have information on the application of MSPC on real cases. Finally academic statisticians are expected to address theoretical perspective on MSPC. The selection of the interviewees started with a collection of possible candidates, mainly locating within the Netherlands. Considering the complexity of the topic, we decided to conduct a face-to-face interview, and the interviewees must be reachable. In addition, having face-to-face communication can lead to more insights during the discussion. Table 3-1 is the list of interviewees. Due to privacy concern, the details of the interviewees and their companies are not provided. Table 3-1 List of interviewees Position of Interviewees Working Environment SPC Consultant Research institute SPC Consultant Research center of a consumer electronic manufacturing company. Industrial Statistician Food manufacturing company Industrial Statistician Paper-making company Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 29 Chapter 3. Industrial Practice Chemometrician Pharmaceutical Company Academic Statistician University 3.2. Objectives of the Interviews To perform an effective interview, we must know what information that we intend to obtain from it. Therefore, we need to formulate the objectives beforehand, and they are listed as bellows. 1). To obtain information of manufacturing industrial companies, regarding the technique of statistical process control. 2). To understand the perspectives and expectations of practitioners regarding the application of MSPC technique. 3). To analyze the potential workability of MSPC technique in industrial plants. 4). To understand the barriers of MSPC implementation in industrial plants. 3.3. Insights from the Interviews Different questionnaires were designed for each target groups, the questionnaires are provided in Appendix D. Basically the questions were designed in a style of open questions, because we are mainly interested in the qualitative information, instead of statistical analysis. The important comments generated from the interviews were screened and listed in the following section. Most of the comments seem straightforward and practical. They cover various aspects, such as technical, economical, organizational and etc. These opinions have broadened our insights on SPC implementation as well as MSPC implementation. The barriers and the niches of MSPC implementation also have been raised. 1. Choosing the right tool and using it correctly is the fundamental to achieve successful SPC. Shewhart control chart is still an effective tool if it is well applied (Statistician of a consumer electronic manufacturing factory). 2. Process system needs to be well investigated. It is in vain to monitor the variables which are not critical to the quality of product. Using the right tool and using it in a correct way both are necessary to achieve the successful result (Statistician of a consumer electronic manufacturing factory). 3. The nature of production shall be considered before implementing MSPC. Certain types of production, for example the chemical process, food production may seem suitable to implement MSPC, because the process is highly complex. Applying MSPC has a better chance to control the process and lower the defect rate before the products are manufactured (Statistician of a consumer electronic manufacturing factory). Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 30 Chapter 3. Industrial Practice 4. However, there are also many other niches for MSPC. It may not be efficient to use MSPC to monitor the process variables if the process is relatively simple. Yet MSPC also can be applied to monitor the product quality. For example if a product has many attributes of quality, these attributes can be monitored with one MSPC control chart instead of numerous Shewhart control charts (Statistician of a Food manufacturing factory). 5. The SPC activities should be categorized into different levels. For instance, shop-floor operator and a R&D center should are working on different levels of sophisticated problem. Utilizing different specializations of the employees is more efficient. MSPC has been applied in our R&D center and it is an effect tool for high level SPC problem (Statistician of a Food manufacturing factory). 6. The interruption to the process due to SPC activity needs to be reduced as much as possible. It would not be surprising that a plant rather continues the process with foreseeable higher defect rate than stops the production process for minor improvement. It also implies that application of MSPC needs to be simple, effective, and efficient. Economical concerns of the entire company can never be put aside. (Statistician of a consumer electronic manufacturing factory). 7. SPC education is necessary. It is very often to see the plant operators overreact on the variance due to common causes. Besides, preparing plant staff the knowledge of SPC may increase the motivation of plant staff and further improve the process performance (Industrial SPC consultant). 8. Reacting correctly when the process shows out-of-control signals is important. Shewhart control chart or MSPC control chart only gives operators a hint that something is wrong in the process. Without investigating the root causes and taking right action, the process does not fix its own problems automatically. (Statistician of a consumer electronic manufacturing factory). 9. Complexity is a major barrier of MSPC implementation. For a long time, MSPC bears the image of complexity, difficult to use, difficult to interpret. Computer aid can efficiently help plant staff to perform the complex calculation and also produce the results in a graphical way which is easier for managers, engineers and operators to communicate in the same language. Thus, computer aid is considered a great catalyst of MSPC implementation. (Industrial SPC consultant). 10. SPC is a necessary approach to achieve higher quality level. However, the company should choose a proper tool depending on several criteria. For instance, the types of the production and the current stage of quality performance. For simple process system which contains less process variables or the variables are quite independent, Shewhart control charst would be sufficient. Also Shewhart control can be a good choice for the company to improve the current quality performance to a certain high Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 31 Chapter 3. Industrial Practice status. For complex process system or aiming at really high quality performance, adopting MSPC can be a proper choice. The company should choose an economical choice, because any implementation takes price. (Industrial SPC consultant). 11. SPC is not simply a statistical issue. To support a company to implement SPC or even MSPC technique involves the organizational aspect as well. In many case, the poor process quality control resulted from a lack of integration. (Industrial SPC consultant). and economic aspects need to be considered. From the view of a 12. Organization company owner, MSPC is valuable when it can generate extra value to the total profit. MSPC is only part of the quality management of a plant. The level of the whole quality management should be improved to a certain level; otherwise, MSPC implementation is not an economic choice. (Statistician of a consumer electronic manufacturing factory). 13. The SPC technique to be implemented is expected to be simple, easy to use and robust. (Statistician of a paper manufacturing factory). 14. Pharmaceutical production requires highly specialized manufacturing process. MSPC in fact is practically adopted to improve the quality of production and it works very well. (Chemometrician from pharmaceutical company). 15. The quality of the final product is determined by the quality of the entire process. So the process needs to be well monitored and controlled from the beginning to the end. Any abnormal variation during the process can lead to defect product in the end and should be avoided as much as possible. (Chemometrician from pharmaceutical company). 3.4. The Gap between Academic Field and Practical Field of Statistical Issues So far we have looked the issue of MSPC implementation from two aspects. Theoretical development was discussed in Chapter 2, and the survey of practical field was performed in the previous section as well. By looking at this issue from both sides, we will study the gap between academic and practical fields, and try to understand how they are disconnected. The result of this analysis can be a good guideline for us to further construct the bridge between these two fields. Table 3-2 is the summary of the gap analysis. Table 3-2 The gap between academic and practical fields. Academic field Practical field Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Potential remedies Page 32 Chapter 3. Industrial Practice Scientific oriented. The theory Difficult to understand what Continuous education and involves too much MSPC is. Therefore, people present the MSPC theory in a mathematics and statistics. hesitate to use the tool that less scientific format can lower they are not familiar with. the reluctance of practitioners. Academic development is still Several different theoretical Complex and ill-defined ongoing. Various approaches approaches are available. knowledge may easily are still being explored and Practitioners may have overwhelm the practitioners’ experimented. difficulty to understand them; motivation. Screening the and they do not know exactly approaches and presenting a which approach to follow. clear structure of the theory overview can be a good start. Also providing a clear guidance and instruction of MSPC application can be a great support. Academic research often What practical field needs is The theoretical part of MSPC concentrates on theoretical an effective, robust tool. It is indeed more complex than exploration. Although the should be easy to understand, the traditional control chart. effectiveness of MSPC is easy to operate and provide However, it can be overcome validated theoretically, the prompt information. by additional support, for feasibility of practical instance, software implementation seems not development, automatic much emphasized. monitoring/alarm system, etc. The implementation of MSPC can be boosted more easily if the complex statistics is supported with computer software. Traditional control chart The investment of (Shewhart control chart) may Implementing MSPC can be be still sufficient. The effort formidable, in terms of finance, and cost to implement a human resource, hardware, complex technique is too high. etc. To maximize the utility of this MSPC, the nature of production should be Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 33 Chapter 3. Industrial Practice investigated in advance. For instance, MSPC may appear to be a proper technique for complex process where variables are highly correlated. Whereas traditional control chart is a more economical choice for a simple process system. In order to have a clear focus, Statistical process control A production plant is absolutely the framework of MSPC (SPC) is just part of the entire not just forwarding input into research is often simplified. quality management work. process and output being The contribution of SPC may produced. In addition to SPC, be limited if other parts of the many other techniques need to process are not improved in be involved. For example, parallel. design of experiment (DoE), automatic process control (APC) are often applied to improve a process system. Proper technique should be well chosen for different situations. The value of SPC or even How to implement the MSPC can reveal only when advanced technique in a its contribution can reflect on proper timing is an important the overall benefit of the concern. For example, a plant company. currently with poor production quality, then traditional statistical process control chart could be already sufficient with respect to the cost of implementation. MSPC is one of the techniques to improve process The quality staffs still often rely on experiences and feelings. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. The implementation of a new technique, such as MSPC, Page 34 Chapter 3. Industrial Practice performance. To implement a Lacking of statistical process requires tremendous technique is different oriented. control knowledge is a barrier investment and very often There are many aspects need to implement MSPC. encounters foreseeable to be taken into account. For resistance. Hierarchical instance, how to set up the organization of quality plan, how to train the management department can employees, estimate the be a good advice. Higher level investment and etc. statistician should have sufficient MSPC knowledge, whereas lower level operators should receive less theoretical but more practical knowledge. Such organization is more efficient in terms of human resource expenditure and the company is also able to perform the internal training to improve the quality of employees. 3.5. Conclusions of the Interviews In this chapter, we performed the interviews with quality people working in the practical fields, and we studied the reasons why the academic development is not implemented into practice. With referring to the objectives that we have set in Chapter 3.2., we would like to make some conclusions. 1). Statistical process control (SPC) is well recognized by industries as an effective tool to achieve higher production quality and traditional control charts, namely Shewhart control charts are often used in practice. Nevertheless, the actual quality improvement sometimes is limited because the quality people do not apply the Shewhart control charts in a correct way. Common mistakes sometimes happen, for example, quality people are confused with control limits of control charts and the specification limits (please refer to Chapter 2.1.), or the variables being monitored are not critical to the quality of final product. This situation shows the education in terms of SPC knowledge, of quality people still needs to be strengthened. The theory of MSPC technique is considered even more complex than Shewhart control charts. Therefore, the education indeed is a necessary efforts, and this is just part of the investment of MSPC implementation. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 35 Chapter 3. Industrial Practice 2). Regarding the workability of MSPC in manufacturing industries, we may conclude that MSPC appears to more effective than Shewhart control charts in the continuous process and batch-wise industries, whereas less effective in discrete process industries. In continuous process and batch-wise process, for instance, chemical industry or plastic injection industry, there are often many variables need to be monitored and controlled during the processing period, therefore MSPC has a greater chance to detect out-of-control events than Shewhart control charts. On the other hand, for discrete process, for example, automotive assembling industry, the quality of final product is generally determined by the quality of the assembling components, and MSPC seems to be little help on quality improvement. Nevertheless, the niches of MSPC application can be created. A nice example was introduced by one of the interviewees. The MSPC is applied to check all the quality characteristics of two types of coffee. More than 30 quality characteristics are monitored, including the color, the taste, the bitterness, the smell, etc. Some of these quality characteristics are correlated, and MSPC can be a very good technique to monitor them. These two types of coffee are produced with different prices of raw materials. So the company may lower the production cost by using the cheaper raw material, if they can successfully make the quality characteristics of the coffee with cheaper material achieving the same level as the one produced with expensive material. In this case, MSPC is not applied during the process but for the quality characteristics of final product. 3). Due to the intention of this research is to facilitate the implementation of MSPC into practice; we also would like to conclude what the expectations are from the view of practical fields. These expectations will be always taken into consideration in the further research development. 3.1). The implementation of MSPC needs to be a gradual, gentle progress. Harsh and sudden extra work may induce great resistance. It is important to consider the shop floor plant people may have difficulty understanding the sophisticated statistical theory. Minimizing the extra work derived from MSPC implementation and present the scientific theory in an easier format can facilitate the process. 3.2). The tool needs to be simple and easy to adopt. In general, Shewhart control charts are well-known in the industrial plants. However, it is found that still very often Shewhart control charts are not correctly applied and lead to little help of process improvement. Obviously MSPC is even more complex than Shewhart control charts, so there must be a clear and simple approach which can step-by-step instruct the plant staff to adopt it correctly during the application. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 36 Chapter 3. Industrial Practice 3.3). Computer software development can be a great support for MSPC implementation. The main resistance from the plant staff is that they need to perform complex calculations to obtain the result, which is over their capacities. Besides, it is also not realistic to request quality people to perform MSPC technique with manual calculation. Most of the plant staffs are familiar with collecting data from the process and to taking action according to out-of-control action plan (OCAP) when it is necessary. To successfully apply MSPC technique, the critical part is something between correctly interpreting the data from MSPC and taking right action. Thus, it is expected to construct the computer software to support plant staff on extracting the information from a large amount of raw data and generating clear instructions for plant staff to follow. 3.4). Graphs are more preferable than numbers and texts. In general the information is easier to be understood with graphs than pages of texts. Besides, for communicating and educating people from different levels, a graphical tool is a better choice and computer software development is also possible to support largely on this requirement. 3.5). The added value from implementation. The most straightforward added value for company is the overall profit growth from the improvement of process performance. To reach this profit growth, the return from the process performance improvement must climb over the investment on the implementation. Therefore, the industrial plants need a program and tool that are able to provide foreseeable improvement of the current situation. 3.6). A profound implementation guideline. This guideline must contain general instructions and practical tools for a production company. Different levels of employees, including management level, R&D specialists and shop floor operators all need to be involved. Each role has to know how to correctly and effectively conduct the implementation work. It is for sure not an easy task for plant to implement a new technique, especially a complex one. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 37 Chapter 4. MSPC Implementation Guideline Chapter 4. MSPC Implementation Guideline After understanding the theoretical development of MSPC and the perspectives from practical field, we are going to construct an implementation guideline so as to fill up the gap and implement MSPC into practice. There will be four elements in MSPC implementation guideline, which are shown in Figure 4-1. The theoretical and practical concerns that we have discovered will be incorporated in this implementation guideline and we will elaborate these four elements in this chapter. MSPC Implementation Guideline 1). Method Model for (M)SPC MSPC Plan 2). MSPC Diagnosis 3). Process Control MSPC Training Team Approach Management Involvement Figure 4-1 MSPC Implementation guideline. 4.1. MSPC Plan We have understood the nature of MSPC is complex and it has been viewed as one of the largest barriers for industrial practitioners. Therefore, a simple and clear instruction will be considered as an effective tool to achieve successful MSPC implementation. We develop two practical tools which are Method model for (M)SPC, and MSPC Diagnosis. The first one is a general instruction for industrial practitioners to know how to choose a proper SPC technique under different circumstances. The second one is the supplementary tool when MSPC is applied. It helps practitioners to identify the problematic variable(s) and correctly react on it when an out-of-control event occurs. 4.1.1. Method Model for (M)SPC There are various types of production. Depending on the nature of the production process, different SPC techniques should be applied in order to achieve the quality improvement economically. Figure 4-2 is the scheme of Method model for (M)SPC, and further details are explained as follows. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 38 Chapter 4. MSPC Implementation Guideline Step 1. Investigate the process system. Understanding the nature of the process system is the first step toward the application of SPC. The practitioners should know what variables need to be stable in order to achieve stable output, and then these variables are suitable to be monitored with control chart. One way to do this is to trace backwards from the final output and screening the possible variables that have influences to the quality of outputs. In this stage, consulting the experienced operators and performing the site investigation can be very helpful. In addition, monitoring a large amount of variables is inefficient in terms of economical concern and effectiveness concern. The critical control points should be identified and the criterion is that the critical control point will lead to significant impact to the quality of output when it goes out-of-control. Step 2. Break down the process system. When the entire process system is too complex or there are too many process steps, breaking down the process system into logical sections, which have highly correlated variables within section but less correlation between sections can be a good advice. Observing the actual production process and discussing with experienced quality people can lead to a general idea of the process system. Applying statistic software to screen the process variables can provide more quantitative information on breaking down the process system. In addition, with smaller monitoring system unit, it is easier to implement (M)SPC, and easier to diagnose the responsible variable(s) when an abnormal measurement occurs. Nevertheless, breaking down the process system should be done in a logical way; otherwise, we may run into a risk of losing information of the process. Step 3. When the number of monitored variables is only 1 (N=1), then it is suggested to use Shewhart control charts. The application of Shewhart charts and relevant information can be found in Chapter 2 and Appendix A. When the number of monitored variables is more than 1 (N>=2), then we need to examine whether these variables correlate with each other. The paired correlation of a group of variables can be examined with generic statistic software (e.g. SPSS, STAGRAPHICS). Correlation r=0.3 can be used as a criterion to decide the strength of the correlation between variables (e.g. if statistically correlation r<0.3, it is considered little or weak association between variables). Step 4. Construct Hotelling’s T2 control chart for the process measurement from a period of in-control process data, when the inversion of covariance matrix is available. If the covariance matrix is nearly singular or not possible to calculate inversion matrix, construct the TA2 control chart and SPE plot for the measurement. If the number of variables is large Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 39 Chapter 4. MSPC Implementation Guideline and they are not correlated between each other, instead of constructing Shewhart control charts for all of them, constructing one Hotelling’s T2 control chart to monitor all variables is also an alternative. There are several computer software packages which support the construction of Shewhart control charts. SPSS, STAGRAPHICS, Conerstone are all workable. For multivariate statistical process control, the software package for constructing the Hotelling’s T2 control chart now is available but not many. STAGRAPHICS (version Plus 5.1) can calculate the Hotelling’s T2 statistic for each observation and also construct the Hotelling’s T2 control chart which is very convenient. Stept 5. In-control data and control limit construction. It is important to purge the preliminary data to obtain an in-control data. This in-control data is established as a norm to monitor the future observation and to see whether it significantly deviates away from the norm. The data purging includes identifying and removing outliers and/or substitute missing data with an estimate. Step 6. So far, at least one particular SPC tool (Shewhart control charts, T2 control chart or TA2 control chart and SPE plot) shall be chosen to monitor the process. An out-of-control situation occurs while using Shewhart control charts, then the responsible variable(s) will reveal easily. While using MSPC control chart, the diagnosis of responsible variables(s) of an out-of-control situation will require more analysis procedures. The detailed will be shown in MSPC diagnosis which will be explained in next section. Systematic Pattern in MSPC It should be noted that in Chapter 2.1.2, we have addressed the issue of systematic pattern in Shewhart control chart and provided Western Electric run rule as a detecting tool. Although it is obvious that systematic pattern in the control chart indicates extraneous sources of process variation, yet very little research addressed the detection of systematic pattern in T2 control chart. For multivariate statistical process control chart, the information of systematic pattern is difficult to interpret because the T2 statistic value is a synthetic information of all the variables. It is a generalized distance between an observation to its sample mean, and it has no clear physical meaning of any one variable in the variable set. Besides, the Western Electric run rules is not an appropriate tool to apply, because T2 statistic has a non-normal distribution. (Mason, Young and Tracy, 2003). We suggested that Shewhart control chart of each variable should be constructed and examined independently when an out-of-control situation occurs in the multivariate statistical process control chart. The procedures of examining systematic pattern of individual variable can be referred to Chapter 2.1.2. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 40 Chapter 4. MSPC Implementation Guideline By now the MSPC has been incorporated into Method model for (M)SPC scheme. Step 1 and step 2 do not really require statistical knowledge. People who actually operate the system are the excellent source to consult to. Apart from that, Hotelling’s T2 control chart is already supported by some computer software, collecting data from the process will be a familiar task for operators as what they do in applying Shewhart control charts. From our survey, inverting covariance matrix is rarely seen and Hotelling’s T2 control chart would be applied in most of the cases. Therefore, the implementation of MSPC does not increase too much extra work for the plant staffs. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 41 Chapter 4. MSPC Implementation Guideline Step 1 Process Legend : investigation Decision point Decision option Step 2 Process breakdown (optional) Step 3 N=1 N: number of variables Action N>=2 No Variables correlated? (r > 0.3) Yes Step 4 Yes No Shewhart control chart N is too many to monitor separately. Yes No Inversion of covariance matrix available? T2 control chart TA2 and SPE control chart Step 5 In-control data and control limits construction In-control data and control limits construction In-control data and control limits construction Step 6 No Continue process. Monitoring future observation. Out-of-control occurs? Yes Responsible variable(s) are found. Investigate the root causes. No Continue process. Monitoring future observation. Out-of-control occurs? Yes Apply MSPC diagnosis. Figure 4-2 Method model for (M)SPC Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 42 Chapter 4. MSPC Implementation Guideline 4.1.2. MSPC Diagnosis MSPC Diagnosis (Figure 4-3) is applicable when an out-of-control situation occurs. Although MSPC control chart is mainly used for monitoring two or more variables which are correlated between each other, it is also a good tool to reduce the work of monitoring many individual Shewhart control charts at a time. Due to the concern of plant staff’s reluctance to accept MSPC, we tried to make the scheme as simplified as possible. The step-by-step instructions for using MSPC diagnosis are described below (also see Chapter 2.3.2. T2 diagnosis with Principal component analysis). Step 1. Compute Normalized PCA Scores (ta/sa) according to the following formula. The eigenvalue and the eigenvector of Principal Component Analysis can be obtained from a general statistical software packages. The rest of the calculation is also possible to perform with Excel spread sheet. n t a = Pa' (x - x) = ∑ pa,j (x j - x j ) j =1 sa = λ a The principal component(s) with higher normalized scores (ta/sa) can be further investigated with contribution plotting (MacGregor, et al. 1994). Step2. Construct overall contribution plot. 1). Select the k normalized scores with high values. Normally the largest two or three normalized scores will be sufficient. 2). Calculate contribution of a variable xj in the normalized score. Cont a,j = ta pa,j (X j − X j ) λa Conta, j is set to zero if it is negative. 3). Sum the total contribution of variable xj. k CONTj = ∑ (cont a,j ) j =1 The overall average contribution per variable generates an overview contribution of each variable in one plot, which is very convenient. Besides, graphical information is easier to understand. The operator can clear see which variables have with higher contribution for a particular out-of-control event and conduct further investigation work. The function of diagnosing the responsible variables from Hotelling’s T2 control chart Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 43 Chapter 4. MSPC Implementation Guideline so far is rarely found in computer software packages. For example, STAGRAPHICS (PLUS 5.1) is one of the softwares supporting multivariate statistical process control but it only supports to compute the Hotelling’s T2 statistic of measurement, no further diagnosis analysis. Legend : MSPC Decision point application Decision option Action Yes No Monitoring future observation. Out-of-control occurs? PCA diagnosis approach Continue process. Compute Normalized PCA Scores (ta/sa). Construct overall contribution plot of the out-of-control observation. Responsible variable(s) are identified. Investigate the root causes. Figure 4-3 MSPC diagnosis. 4.1.3. Process control It should be aware that the function of process control chart, either Shewhart control chart of multivariate statistical process control chart, only monitors the behavior the variables. So we will always have two consequences when process control chart is applied. First, let the process continue when there is no signal from the control chart. Second, when the control chart signals, responsible variables await to be further identified. It should be noted that identifying responsible variables does not mean the root causes of an out-of-control situation are located. Without identifying the root causes and taking necessary steps to bring the process back to normal status, Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 44 Chapter 4. MSPC Implementation Guideline the process does not fix itself automatically. Therefore, we need additional tools that can systematically help us find the potential root causes when out-of-control situation is encountered. Cause-and-Effect Diagram (Figure 4-4) is a recommended tool to uncover the potential root causes of an undesired problem. The procedures of constructing a cause-and-effect diagram are summarized as follows (Modified from Montgomery, 1985). 1). Define the problem or the effect that needs to be analyzed. 2). Uncover potential causes through brainstorming. 3). Specify the major potential cause categories and join them as boxes connected to the center line. 4). Identify the possible causes of each major potential cause. 5). Rank the causes to identify those that seem most likely to affect the problem. 6). Take corrective action. The possible reasons and the remedies of an out-of-control situation shall be collected and documented as an Out-of-Control-Action plan (OCAP). It should contain all the diagnostic knowledge and all the operators can follow the standardized procedures to bring the process back to in-control status. Machines Materials Methods Wrong procedure Wrong tool Defect supplier Damaged handling Insufficient warm-up Wrong planning Problem or undesired effect Wrong specification Inexperience High humidity High temperature Faulty gauge Measurement Poor attitude Manpower Environment Figure 4-4 An example of Causal-and-effect diagram. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 45 Chapter 4. MSPC Implementation Guideline Shewhart/Deming wheel (PDCA) The Shewhart wheel is one of the well-known strategies to achieve the process improvement. It is known as PDCA cycle, where P stands for Plan, D stands for Do, C stands for Check and A stands for Act. The idea of PDCA was originated also by Dr. Shewhart, and later popularized by Dr. Deming. The book Out of the crisis (1986), by Dr. Deming is a recommendable reference for further study. Figure 4-5 The Shewhart/Deming wheel (PDCA). Here we summarized the key idea of PDCA. Plan: Identify the problem and possible causes. Cause-and-Effect Diagram can be an effective tool to accomplish this task. Do: Make changes that designed to correct to problem or improve the current process situation. Check: Study the result of these changes that have been taken. Act: Standardize the changes if the result is successful, and document these changes into OCAP. Whereas, search new strategy to improve the process if the changes did not make much progress. Combing the activities that involved in MSPC technique, such as monitoring the process, diagnosing the responsible variables for out-of-control measurement and PDCA strategy, it forms a continuous work flow to continuously improve the process performance. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 46 Chapter 4. 4.2. MSPC Implementation Guideline MSPC Training Taking into account the concerns of lacking of knowledgeable employees and high implementation cost that raised from the interview of practical filed. We recommend different levels of training should be developed to train the employees. First, advanced level of MSPC education should be provided to quality experts in the R&D department. In the beginning of the implementation, the quality experts must be familiar and execute all the steps mentioned in the MSPC Plan. MSPC Plan is a generic instruction, the quality experts should apply and adjust (if it is necessary) the MSPC Plan to fit the real situation. After construction the detailed procedures, they may assign the tasks down to lower level quality people. Second, basic level of MSPC education should be provided to all the shop floor operators, process engineers in the plant. It is necessary for them to be familiar with concept of MSPC and correctly conduct all the relevant tasks assigned from quality experts. The lower level employees should be able to properly collect data, and correctly react to out-of-control event according to Out-of-Control-Action Plan (OCAP). Third, moderate level course for management employees should be provided. In order to avoid the communication gap, it may not be necessary for management employees to perform the statistical analysis, yet they should be familiar with all the procedures happening during the process monitoring. 4.3. Team Approach MSPC Plan can be viewed as the tool to conduct MSPC monitoring, and MSPC Training makes the users understand how to correctly use this tool. During the MSPC Training, we generally define three different levels of users, advanced level (quality experts in research center), basic level (plant people) and moderate level (management employees). The MSPC implementation work is directly conducted by quality experts and plant people in a way of team work. Team approach describes the work scopes of quality experts and plant people and how they should conduct the work by using the skills that they have been trained in the previous step. The guideline of work scope is described as follows. Quality experts: Develop/adjust the MSPC Plan to fit the real situation. Construct OCAP. Conduct internal training, to plant people and management employees. Supervise the work of plant people. Provide expertise to plant people on complex issue. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 47 Chapter 4. MSPC Implementation Guideline Plant people: 4.4. Implement the MSPC Plan that defined by quality experts. Resolve out-of-control situation according to OCAP. Support quality experts on maintaining OCAP. Management Involvement Management commitment is crucial to the success of implementation work. The progress of the entire MSPC implementation should be monitored by management employees. A regular review system and audit system should be established. The management employees not only monitor the process performance, but also continuously drive the plant people to improve their work. On the other hand, except driving force, encouragement is also necessary. Motivate the plant people with rewards, in terms of promotion, bonus, etc., and create the quality culture inside the organization can be helpful. In addition to monitoring the technical issues, management employees should also coordinate with relevant department, such as finance, procurement, marketing, etc. to achieve an overall improvement. For example, when the poor quality results from the inferior raw material, then the task is expended to procurement and finance aspects as well. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 48 Chapter 5. Case Study Chapter 5. Case Study In this chapter, a real case from a paper-making company will be studied. Two main analyses will be conducted. First, we will implement multivariate statistical process control in this case, and validate the approaches that we have developed in chapter 4 (Method model for MSPC and MSPC diagnosis). Second, we will investigate the mechanism of the process system with multiple regression analysis and provide practical recommendations in order to improve the process performance. 5.1. Case Briefing The case for further study is a process unit from a paper-making process. This process unit can be described as follows. A flow containing fibers (named Thickstock) is mixed with another flow containing fillers and water (named Filler). Thickstock will be called Thinstock when diluted with water. So the Thinstock is the mixture of fibers, fillers and water. This dilution with what is called Whitewater to control the solid content in the Thinstock. So the Whitewater increases when the solids percentage in the Thinstock is too high, whereas it decreases when the solid percentage is too low. In addition, a chemical (named Ret) is added before the Headbox to accelerate the separation of solids and the water during the dewatering. Major part of the solids will form the paper at the end, and the liquid leaving the dewatering machine is mainly recycled as Whitewater to dilute the Thinstock again. Finally the mixture flow (named Mixture-in) being pumped into Headbox contains fibers, fillers, water, and chemical. There are two variables are measured in the Headbox, one is the amount of solids of the Mixture-in (named HBtotal) and the other is the portion of the solids due to fillers (named HBash). Also two variables are measured in the Whitewater, one is the amount of solids (name WWtotal) and other is the portion of solid due to fillers (named WWash). The amount of Ret is controlled by measuring the difference between HBtotal and WWtotal. After dewatering, the solid material will form the paper and the weight of paper is measured (named BW). Part of the BW comes from fillers, this portion is named Paperash. Paperash is a way of reducing the production cost, because fibers are more expensive. Within our analysis scope, stable and correct BW value is the target. BW is also a precondition for further processing. A conceptual scheme is provided in Figure 5-1. It should be noted that a certain level of simplification has been embedded, because to fully reflect the real situation only when the entire process system is considered. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 49 Chapter 5. Case Study Figure 5-1 Scheme of paper-making process unit. All the process variables are tabulated as follows. Table 5-1 Process variables. Variables Measuring unit Description Thickstock [ton/hr] Inflow of Thickstock. Filler [ltr/min] Inflow of the fillers and water. HBtotal [g/ltr] Amount of solid in the Thinstock. HBash [%] Amount of solid in the Thinstock due to fillers. Ret [ltr/hr] Chemicals WWtotal [g/ltr] Amount of solid in the whitewater. WWash [%] Amount of solid in the whitewater due to fillers Paperash [%] Amount of paper weight due to fillers. BW [g/m2] Paper average weight. 5.2. MSPC Implementation Before applying the MSPC to this case study, the data needs to be examined. The raw data was provided from the plant supervisor and all the measurements of the process variables are measured by sensors with an interval of one minute. The raw data set is a period of 1001 measurements. To obtain a rough idea what the process behavior is of the variable, we construct the I-chart for each variable with raw material. Surprisingly, we see the control limits (with 3 sigma) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 50 Chapter 5. Case Study are extremely small and numerous of measurement fall outside of the control limits. The reason is that, the sample size of measurement is only one (that is way we used I-chart), and the control limits of I-chart were calculated based on the moving range of consecutive measurements. Due to the small interval (one minute), the moving range between every two measurements is extremely small. We recognized that the control chart with such small measuring interval is meaningless, because it does not reflect the real process behavior. After some trials, we construct the control chart for each variable with an interval of 20 minutes. Therefore, we have a preliminary data set containing 51 measurements and we will continue the analysis with this data set. Measurement Interval of I-chart Here we will elaborate why the one-minute interval is not an appropriate choice. The idea of the control limits in the Shewhart control charts is to represent the process behavior, which means the control limits should correspond to the variation of the process. A graphical illustration will be easier to understand. For upcoming explanation, we will take variable HBtotal as an example. Figure 5-2 is the I-chart of HBtotal that constructed from 1001 measurements with one-minute interval. The control limits are very close to centerline, and they are directly influenced by the small moving range of consecutive measurements. Control Chart: HBtotal 14.8 HBtotal UCL = 14.5229 Average = 14.4954 LCL = 14.4679 14.6 14.4 14.2 14.0 989 963 937 911 885 859 833 807 781 755 729 703 677 651 625 599 573 547 521 495 469 443 417 391 365 339 313 287 261 235 209 183 157 131 105 79 53 27 1 Figure 5-2 I-chart of HBtotal with one-minute measurement interval. With such small measurement interval (one-minute), the measurements are severely auto-correlated. Auto-correlation means that the value of each measurement is dependent on the previous one. Therefore, while using control charts, the assumption that the data from the process is normally and independently distributed with its mean and standard deviation is violated. By selecting the measurements with larger interval can lower the tensity of auto-correlation situation. The control limits of I-chart will become wider as we choose larger interval, because the Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 51 Chapter 5. Case Study moving range between measurements also increases. Now we come to a question, to what measurement interval is the proper choice? The percentiles of histogram plot can be a good reference. Now, if we plot the histogram of all these measurements (Figure 5-3), we see that the variance of the process is much wider than the control limits (14.47, 14.52) in Figure 5-2. The control limits of 3-sigmas away from the mean in the histogram plot are approximately (14.10, 14.89). Histogram from Dataset 240 220 200 180 160 140 Count 120 LEGEND Normal (14.4954,0.132827) Dataset 100 HBtotal Count 80 60 40 20 0 14 14.05 14.1 14.15 14.2 14.25 14.3 14.35 14.4 14.45 14.5 14.55 14.6 14.65 14.7 14.75 14.8 HBtotal Figure 5-3 Histogram of HBtotal measurements. When we adopt 20-minute interval, the control limits of I-chart has become (14.2, 14.8) and 51 measurements remains. Figure 5-4 is the I-chart of HBtotal with 20-minute interval and it is considered acceptable for further analysis. Control Chart: HBtotal 14.8 HBtotal UCL = 14.7668 Average = 14.4848 LCL = 14.2028 14.6 14.4 14.2 14.0 51 49 47 45 43 41 39 37 35 33 31 29 27 25 23 21 19 17 15 13 11 9 5 7 1 3 Figure 5-4 I-chart of HBtotal with 20-minute interval. During this case study, we intend to apply the approaches developed in Chapter 4 (Method model for MSPC and MSPC Diagnosis) so as to validate the effectiveness of these tools. Therefore, in the following section, we will perform the analysis step by step associated with detailed explanation and also provide a clear overview by highlighting the decision path in Figure 5-5. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 52 Chapter 5. Case Study Step 1 Process Legend : investigation Decision point Decision option Step 2 Process breakdown (optional) Step 3 N=1 N: number of variables Action N>=2 No Variables correlated? (r > 0.3) Yes Step 4 Yes No Shewhart control chart N is too many to monitor separately. Yes No Inversion of covariance matrix available? T2 control chart TA2 and SPE control chart Step 5 In-control data and control limits construction In-control data and control limits construction In-control data and control limits construction Step 6 No Continue process. Monitoring future observation. Out-of-control occurs? Yes Responsible variable(s) are found. Investigate the root causes. No Continue process. Monitoring future observation. Out-of-control occurs? Yes Apply MSPC diagnosis. Figure 5-5 MSPC decision path of case study. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 53 Chapter 5. Case Study Step 1. Process investigation. The description of the process unit and its scheme has been provided in the section of case briefing. As we mentioned in Chapter 4, monitoring a large amount of variables is not efficient. Only the critical quality characteristics should be selected and monitored. Another concern here is that the process unit contains several loop controlling systems, therefore many variables change during processing due to automatic controller. What we need to do is to search which variables need to be stable to achieve a stable BW. If we look at the process scheme, we will see that the HBtotal and HBash are the quality characteristics of the input (Mixture-in) flowing into the dewatering machine. Therefore, when the processing inside the dewatering machine is stable; the output (BW) should be stable if the input is stable. So HBtotal and HBash are two variables to be monitored with statistical control chart. As mentioned in the case briefing, the Ret is automatically controlled due to HBtotal and WWtotal. Thickstock and the Whitewater inflow are also automatically controlled to maintain stable quality of the Mixture-in. Only the Filler is a fixed flow and it can be monitored with statistical control chart. Therefore, we will perform the multivariate statistical process control for these three parameters, Filler, HBtotal and HBash. Step 2. Process breakdown. After conducting the preliminary investigation in step 1, we see no reason for further decomposition. We continue the analysis according to following steps. Step 3. Number of variables. Filler, HBtotal and HBash are the variables to be monitored, so it is clear the number of variables N is three. The next action is to examine the variable dependency. The correlation between variables is a good indicator telling us how intense the variables are related to each other. Table 5-1 is the variable correlation generated from a period of process containing 51 measurements. Here we see that there is a moderate positive correlation between HBtotal and HBash, so this is a good condition (but not absolutely necessary) to adopt multivariate statistical process control. Table 5-1 Correlation of process variables Correlation Filler HBtotal HBash Filler 1.00 -0.17 -0.19 HBtotal -0.17 1.00 0.58 HBash -0.19 0.58 1.00 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 54 Chapter 5. Case Study Step 4. Checking the inversion of covariance of variables and choosing T2 control chart. Normally it will be problematic when the number of variables is really large. In this case dealing with three variables, the inversion of covariance is not a problem (it has been further examined with statistical software). So it is applicable to adopt T2 control chart. Step 5. In-control data and control limits construction. A set of data containing 51 measurements with an interval of 20 minutes was taken from a generally in-process period. The data was further analyzed by I-chart (Individual chart) with customary plus/minus 3 sigma control limits and the problematic measurements were removed. After that, we obtained a sample set containing 21 measurements. The T2 control chart was also constructed (see Figure 5-6) to see whether any observation containing a problematic relationship between parameters. In this case, no indication of out-of-control showed. The measurements (No.1 to 21) are tabulated in Table 5-2. This data is used as a norm to monitor future observation and to further analyze the reasons of out-of-controls, if any. Figure 5-6 T2 control chart of in-control measurements. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 55 Chapter 5. Case Study Table 5-2 In-control process data No. Filler HBtotal HBash No. Filler HBtotal HBash 1 25.432 14.497 3.184 16 24.485 14.508 3.181 2 25.753 14.517 3.172 17 24.762 14.420 3.157 3 25.178 14.696 3.199 18 25.161 14.528 3.196 4 25.318 14.581 3.207 19 25.393 14.580 3.190 5 25.321 14.548 3.205 20 24.773 14.560 3.182 6 24.636 14.596 3.204 21 24.741 14.413 3.174 7 25.055 14.490 3.197 22 24.926 14.480 3.124 8 24.963 14.594 3.182 23 25.661 14.452 3.132 9 24.720 14.528 3.167 24 24.839 14.752 3.275 10 24.347 14.516 3.165 25 24.474 14.503 3.266 11 25.000 14.552 3.162 26 24.598 14.478 3.226 12 24.953 14.401 3.160 27 24.777 14.314 3.191 13 25.269 14.501 3.150 28 24.924 14.366 3.190 14 24.862 14.483 3.165 29 24.478 14.665 3.186 15 25.107 14.513 3.180 Step 6. Monitoring future observation. A period of future observation with eight measurements (No.22 to 29) will be analyzed with T2 control chart and see it if any observation is out-of-control. The Hotelling’s T2 statistic is calculated for each new observation based on the mean and the covariance matrix obtained from the in-control data set. The control limit is chosen with Type I error α = 0.05. The T2 control chart (Figure 5-7) shows the observations 22, 23, 24, 25, 26, 27, and 28 are out-of-control. The control chart signaled us that something went wrong during the observation 22 to 28, yet we do not know which variable or set of variables is responsible for it. So we need to identify those variables with MSPC diagnosis. Figure 5-7 T2 control chart for future observations. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 56 Chapter 5. Case Study MSPC Diagnosis We will apply the approach T2 diagnosis with Principal component analysis (PCA) (details please refer to Chapter 2.3.2). The generic steps can be repeated briefly once more. Normalized PCA scores (ta/sa) are calculated and see which one(s) has/have higher scores. Figure 5-8 shows the chart of overall average contribution per variable is constructed based on the selected high score Normalized PCA. The contribution of each variable to the out-of-control measurement will be shown in this chart, and it gives us the information what problematic variables are. It should be noted that it is suggested to perform the diagnosis approach with standardized data set. Due to the different measuring scale of each variables, the variable with smaller measuring scale will have relatively less weight than the one with larger measuring scale. For a particular observation, each bar indicates the contribution of one variable. The bars represent Fillers, HBtotal and HBash (from the left to the right). Variable contribution 22 23 24 25 26 27 28 Figure 5-8 Overall average contribution of every variable from observation 22 to 28. 5.3. Result of MSPC Implementation We have known the out-of-control observations detected by T2 control chart, and obtained the overall contribution plot which tells us what problematic variables are for each observation. Now we are going to analyze each of the out-of-control measurements and draw conclusions. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 57 Chapter 5. Case Study Table 5-3 The result of MSPC diagnosis. Observation Signaled by MSPC Potential problematic Status variable(s) Signaled BW is by USPC Out-of-control? 22 Yes HBash Exceeding control limit. Yes 23 Yes HBash Exceeding control limit. Yes Fillers Within control limit No HBtotal Exceeding control limit. Yes HBash Exceeding control limit. Yes 24 Yes No No No 25 Yes HBash Exceeding control limit. Yes Yes 26 Yes HBash Exceeding control limit. Yes Yes 27 Yes HBtotal Exceeding control limit. Yes Yes 28 Yes HBtotal Within control limit No !! Yes From the summary in Table 5-3, we can see that from observation 22 to 27, at least one of the three variables exceeded control limits of the I-chart, which means these observations would also receive an out-of-control signal by using Shewhart control chart. In observation 28, there is no any variable falling outside of control limit, but T2 control chart signaled. For the convenience of comparison, the I-chart of three monitoring variables (Filler, HBtotal, HBash) and BW are constructed in Figure 5-9. If we look at these three variables carefully in the I-chart respectively, we will see that in observation 28, the HBtotal has low value while HBash is relatively high. This behavior contradict to the in-control process behavior that we already found that there is a positive correlation between HBtotal and HBash. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 58 Chapter 5. Case Study Figure 5-9 I-chart of Filler, HBtotal, HBash and BW. To provide a better picture, we plot the ellipse control chart (Figure 5-10) just for variable HBtotal and HBash, to see how observation 28 is signaled as out-of-control. The correlation between HBtotal and HBash has been examined. It is a moderate positive correlation, so the control region (the ellipse) is slightly fat. In Figure 5-10, we can clear see that HBtotal and HBash are both within individual control limits, which is traditionally statistical process control approach. But due to the anti-correlation, it falls outside of the elliptical control region, which is signaled by multivariate statistical process control approach. Another point to be discussed is that, in Table 5-3; we compared the status of output (BW) and the variables being monitored. They do not fully correspond to each other. Observations 25 to 28, MSPC signaled, and BW appeared to be out-of-control. Yet, observation 22 to 24, MSPC signaled but BW was actually in-control. The explanation could be that the variables that we have Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 59 Chapter 5. Case Study monitored do not completely represent all the characteristics of the Mixture-in. HBtotal and HBash only measure the amount of solid content and the portion due to fillers. Obvious the chemical content, other possible characteristics, for instance, the viscosity, ph value, purity, temperature, etc. of the Mixture-in are not monitored. Moreover, the processing inside the dewatering machine has been assumed to be a constant state. Due to such circumstance, we see the need of an integrated monitoring activity which can reflect the whole process better. Ellipse control chart 3.28 3.25 28 HBash 3.22 3.19 3.16 3.13 3.10 14.3 14.4 14.5 14.6 14.7 14.8 HBtotal Figure 5-10 Ellipse control chart of HBtotal and HBash. 5.4. Reflection From the findings of this case study, several points can be concluded. 1). MSPC is a more sensitive technique than Schewhart control charts in terms of detecting power. It monitors not only the deviation from its mean of a variable, also monitors the relation between variables. The traditional control chart (e.g. Shewhart control chart) only signals when the deviation of a variable is abnormal. 2). As we have mentioned in Chapter 2, the idea of SPC is to detect the variability of a process due to special causes and improve the process performance by eliminating these causes. In this sense, high sensitive SPC technique such as MSPC is more powerful on detecting the occurrence of special causes than Shewhart control chart. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 60 Chapter 5. Case Study 3). For the case that the variables are highly correlated with each other, the chance that the variables are in-control of its individual control chart but fall outside of the MSPC control limit will be much larger. Under such circumstance, using Shewhart is very likely to miss the timing to signal when the process goes wrong. Therefore signaling out-of-control observation effectively by MSPC helps to reduce the cost on producing defect product. 4). Modern production system is very complex. The example of this case study already has shown that even a small unit of the entire process may contain several control loops, automatic control devices, etc., which make the process more complicated and dynamic. Although statistical process control chart (either Shewhart control chart or Multivariate statistical process control chart) is just one of the process control techniques. In order to successfully deal with a complex process system and improve the process performance, additional integrated technique is definitely required. 5.5. Process Performance Improvement In this section, an example will be provided to show how other technique is applied to achieve better process performance. As already mentioned the process is quite complex, thus, for the ease of demonstration, we will depict a smaller sub-process as an example for further analysis (Figure 5-11). In this sub-process, BW is considered as the output that needs to achieve a certain target value. The information about the input contains HBtotal, HBash and Ret. By investigating and understanding the mechanism between input and output, we may discover more alternative ways to control the process and achieve the desired quality of output Figure 5-11 Sub-process. A multiple regression model will be performed with a period of data set (the data set can be found in Appendix E) and help us to understand the mechanism between inputs and output. We can Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 61 Chapter 5. Case Study express the output as a function of inputs. It should be noted that non-liner relation, such as quadratic term or interaction term can also exist between input and output. So we will experiment the multiple regression with different types of model including linear model, linear with interaction terms and quadratic model. Each of the regression models will be examined statistically. The result is given in Table 5-4, detailed information about multiple regression analysis can be found in Appendix D. Table 5-4 Result of regression model of sub-process. Regression model BW = f (HBtotal, HBash, Ret) Quadratic model BW = 0.03 + 0.34HBtotal + 0.04HBash + 0.79Ret + 0.016HBtotal – 0.19 HBash (Scaled data) (Interaction term is not significant, so it is not included) Quadratic model BW = 378.54 – 165.06HBtotal + 583.36HBash + 0.67Ret + 5.77HBtotal – 91.41 HBash (Original data) (Interaction term is not significant, so it is not included) Adjusted R 2 2 2 2 2 0.70 P r e d i c te d B W G ra p h 3 .1 8 7 4 5 1 4. 48 48 4 3 .5 4 2 1 1 57 B W 15 5.44 8 + /-0 .2 8 5 4 9 7 1 56 1 55 1 54 3 .1 5 3 .2 3 .2 5 H B as h 1 4 .2 1 4. 4 1 4. 6 H B t o ta l 43 Regression model BW = f (HBtotal, HBash, Ret) Linear with interaction term BW = 0.22HBtotal + 0.76Ret model (Interaction term is not significant, so it is not included) (Scaled data) Linear with interaction term BW = 110.97 – 1.13HBtotal + 0.64Ret model BW = 0.22HBtotal + 0.76Ret (Original data) 44 R et 45 (Interaction term is not significant, so it is not included) Adjusted R 2 0.66 P re dicte d BW Gra ph 1 4 .4 8 4 8 4 3 .5 4 2 1 1 57 BW 155.427 1 56 +/-0. 171292 1 55 1 4. 2 14.4 H B to tal 14.6 43 44 R et 45 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 62 Chapter 5. Case Study From the analysis result in Table 5-4, we see that a quadratic model explains the relations of inputs/output to a satisfactory level (Adj-R2=0.70). The regression model was performed with both original and scaled data. Using scaled data, the coefficient of each term in the model is considered as the weight of the term, so we can see which term has stronger influence to the output. It is clear that Ret and HBtotal dominate the response variable (BW) most. In the predicted BW graph, we can see the relation between BW and each input variable. Due to the quadratic model term, HBtotal and HBash both have a curve regression line. Particularly in HBtotal, we found it difficult to interpret the meaning of its relation with BW. The weight of the paper (BW) mainly comes from the solid in the input and the graph shows either increase or decrease of HBtotal can raise BW, which does not reflect the real situation. It should be noted that sometimes a particular regression model may show a satisfactory statistical result, yet it does not correspond to the reality. We should be careful on such situation and check whether there are other types of model can lead to better explanation. Therefore, we continued to try linear model with interaction term. The result is given in Table 5-4, and the positive linear relations reflect our expectation better. Again we see that BW is mainly dominated by Ret and HBtotal with positive relations, whereas HBash has no significant relation with BW. Knowing the relation between variables, we can further control the process in a more predictable way. Figure 5-12 gives us even clear information. To increase HBtotal and Ret with different combination can reach a particular target BW value. An example is given in Table 5-5. The possible recipe of HBtotal and Ret can be decided due to different concerns, such, cost, availability and etc., to achieve an efficient purpose. Predicted BW Contour Pl ot 45 156.2 156 55.8 Ret 44 1 56.8 156.6 156.4 1 56 .2 156 156.2 1 56 155.6 155.8 55.4 155.6 1 55.8 155.2 155.4 1 55 .6 155 43 156.4 154.6 14.1 14.2 Predicted BW 155.2 155 1 54 .8 1 4. 3 LEGEND 155.4 155.2 154.8 156 155 14.4 14.5 14.6 14.7 HBtot al Figure 5-12 Contour plot of HBtotal and Ret with respect to BW. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 63 Chapter 5. Case Study Table 5-5 Possible recipe of HBtotal and Ret Desired BW Estimated BW [g/m2] HBtotal [g/ltr] Ret [ltr/hr] 155.0 155.5 15.0 43.0 155.0 155.4 20.0 34.0 137.0 137.0 6.0 30.0 137.0 137.4 12.0 20.0 However, it has been noticed that the Adj-R2 is not really high (0.66), in the linear model, which means the relation was not explained by the model very well. The possible reasons could be the same as we have mentioned in Chapter 5.3. The characteristics of inputs may not be fully measured, so the current characteristics of the input are not able to reflect the mechanism inside the process. Another possibility is that the measurement error exists. Nevertheless, this example demonstrated how other technique helps to achieve a better process performance. The same exercise can be extended to other parts. For example, consider HBtotal is an output of another sub-process, investigating what the relations are between the output and its inputs. A target value of HBtotal can also be reached by controlling the inputs in a logical way. Once the whole process is well investigated, the plant staff will have better understanding of the process mechanism and be able to improve the process. To close up this section, we can conclude that to achieve the goal of improving the production quality of a particular industry, investigating and understanding the current process system are very necessary and very fundamental tasks. Without such knowledge, further technique application, such as SPC control charts or multiple regression models will have very limited contribution. The idea is that these techniques are applied to analyze and to reflect what actually is happening during the process and after that, we can have greater opportunity to control the process and further improve the production quality. Nevertheless, correct and complete understanding of the mechanism of the entire process prevents our judgment from misled by the incorrect analytical results. Two advices can be given. First, look at the entire production process with broader perspective to have a thorough picture. This information is very useful when breaking down the process into smaller units for detailed analysis and also very helpful to make the right judgment on the results of analysis. Recall the multiple regression model (the quadratic model), the result shows the BW would increase when HBtotal either increases or decreases, which is not true. Understanding the process and having logic thinking help us to make the correct judgment. Second, start with smaller part of the process while conducting detailed analysis. By doing so, it is easier to find out what really happens inside the process. Then we may gradually expend the scale of process unit for analysis. For instance, if we apply the multiple regression with larger scale process unit immediately, which includes too many variables in it, we may have problem to identify the causal relation between dependent variable and independent variable. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 64 Chapter 6. Conclusions and Recommendations Chapter 6. Conclusions and Recommendations During this research, we began with the importance of the quality and how statistical process control (SPC) was developed. Then we addressed the Shewhart control charts and the multivariate statistical process control (MSPC). Due to the gap between academic development of MSPC and its practical application, we started to investigate the situations of both sides so as to further facilitate the implementation of MSPC and support the industrial plants to utilize the benefit of MSPC technique. In this chapter, we will summarize the main findings of this research and provide several recommendations. At the end, a prospect of future research will be suggested. 6.1. Conclusions The objective of this thesis research is to make recommendations for implementing multivariate statistical process control (MSPC) in a process-industry plant by providing clear interpretations of MSPC and suggestions to quality management staff in the plant. We will evaluate the achievement of this research work by reviewing the research question and its sub-questions. The main research question is defined as, “What are the difficulties of multivariate statistical process control (MSPC) implementation and how quality management staff can be supported to facilitate MSPC in a process industry plant?” There are five sub-questions, which are defined as: 1). What is the essence of MSPC? 2). What are the expectations from quality management staff in the process industry plant? 3). What tools can be provided to make the interpretation of MSPC results easier? 4). What advices can be provided to cope with the time axis problem for MSPC? 5). What recommendations can be provided to quality management staff in the process industrial plants? The findings and the contributions of the research which provide answers to all the sub-questions are concluded as follows. Clear Introduction of Statistical Process Control. Although multivariate statistical process control (MSPC) is the main topic of this research, various types of Shewhart control charts and its application were also addressed (in Chapter 2.1) so as to make this study more complete. In addition, with the knowledge of Shewhart control charts, it is easier to further understand MSPC. Discussion of MSPC. In Chapter 2, Hotelling’s T2 statistics was well elaborated. We began with simplest case, bivariate process control chart to illustrate to idea of MSPC. After that, a generic MSPC control chart for Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 65 Chapter 6. Conclusions and Recommendations multiple variables was extended. Following the explanation, a practitioner will be able to perform the multivariate statistical process control chart to examine whether a period of process is in-control. Apart from typical Hotelling’s T2 statistics, an adjusted type - T2A & SPE plot - was discussed. This method can be applied in case the inversion of covariance matrix is not available (see Chapter 2.2.2.). To overcome the difficult interpretation of MSPC, two approaches - MYT T2 decomposition and T2 diagnosis with Principal component analysis (PCA) were introduced with a case demonstration. In the end of Chapter 2, we provide a clear overview of multivariate statistical process control approaches (Figure 2-14) and comparison of these approaches. Understand the Perspective and the Need from the Practical Field. In order to facilitate the implementation of MSPC into the practical field, we also investigated the perspective from practical field so as to understand the gap between the academic and practical fields. The investigation was summarized in Table 3-1, and the potential remedies were generated as well. Several conclusions have been made and incorporated during the MSPC Implementation Guideline in Chapter 4. For example, the tools should be simple, easy to follow, graphical format, associated with available software is preferable. Other comments will be further elaborated as part of the general recommendations in the next section. Development of MSPC Implementation Guideline By knowing the situations of both academic progress and the practical field, we have developed MSPC Implementation Guideline. This guideline contains four elements, which are MSPC Plan, MSPC Training, Team Approach, and Management Involvement. Especially in MSPC Plan, two practical tools were constructed, which are Method Model of (M)SPC and MSPC Diagnosis. Method model of (M)SPC is a decision flow chart, which supports the practitioners to apply the proper statistical process control chart for different circumstances. It covers the situations of using Shewhart control charts and using MSPC control chart. MSPC Diagnosis is designed to interpret the result of MSPC. Because the MSPC only signals the occurrence of an out-of-control event, it does not provide further information about what the problematic variable(s) are. MSPC Diagnosis basically describes the key steps of T2 diagnosis with Principal component analysis (PCA) introduced in Chapter 2. As already mentioned, one of the drawbacks of MSPC is that it does not tell us what the problematic variable(s) are. We have screened two advanced approaches, and the reason why we propose T2 with PCA approach has been addressed in Chapter 2.4 (approach discussion). So far, some statistical process control software packages provide the function of calculating Hotelling’s T2 statistics, yet the function of MSPC diagnosis is rarely found. MSPC diagnosis can be used as a conceptual specification for software programming in the next stage. Getting started is important. Practitioners may be overwhelmed with various scientific knowledge, approaches, etc., although they all can be workable. Providing a Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 66 Chapter 6. Conclusions and Recommendations clear and easier guidance on selecting the approaches indeed can save plenty of time and avoid unnecessary confusion. The other parts of MSPC Implementation Guideline emphasize how to practically implement this technique into industrial plants and the concerns of financial, organizational, managerial aspects were incorporated as well. The MSPC Implementation Guideline is a practical tool to facilitate the MSPC application in industrial plants. Systematic Pattern of MSPC. This problem and possible solution have been addressed in Chapter 4.1. It is suggested that Shewhart control chart of each variable should be constructed and examined independently when an out-of-control situation occurs in the multivariate statistical process control chart. The procedures of examining systematic pattern of individual variable can be referred to Chapter 2.1.2. Learning from the Case Study. Although MSPC advocates often emphasize the effectiveness of this advanced technique, the real case application and implementation is rarely found. It is understandable that to have a clear focus and framework for a particular research, a certain simplification is inevitable. However, it is still important to step backwards and look at the problem in a broader view. Several points are elaborated as follows. a. MSPC control chart is examined as an effective technique in the case study; even the correlation of variables is not really high. It is predictable that in the case where the number of variables is large and the correlation between them is high, then using MSPC would become much more valuable than using Shewhart control charts. Because the probability that an out-of-control event falls outside of MSPC control limit but within Shewhart control chart separately will become much larger b. The real production process is often a very complex system. It may contain various control devices, such as, automatic controller, automatic sensors, etc. Control chart (either Shewhart control charts or MSPC control chart) is a necessary (but not the only) tool to conduct statistical process control and further achieve higher process performance. Using the proper tool at right place is important; for instance, using a control chart to monitor the fuel consumption of a reactor where the temperature needs to be stable is meaningless. Temperature is the one needs to be monitored! c. To improve a complex production process, many techniques are required. For example, Design of experiment (DoE), (Non)-Linear multiple regression analysis, and SPC control charts are often applied. An integration of all these technique may utilize their effectiveness to a higher level. d. Process investigation is crucial. Academic papers often start to talk about MSPC analysis immediately with “given” variables. In reality, without investigating the process Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 67 Chapter 6. Conclusions and Recommendations system, we may run into the risk of monitoring “fuel” instead of “temperature”. Learning from doing or communicating with experienced experts can be helpful because process investigation involves large amount of tacit knowledge and experience, rather than purely mathematics or statistics knowledge. 6.2. Recommendations Apart from the findings and developments that have been concluded, it is necessary to address several points in this section. These points are closely relevant to this research, nevertheless they are rather conceptual or managerial oriented. Financial Evaluation. The benefit will not fall off from the sky. The implementation of statistical process control is a huge investment in terms of human resource, equipment, finance, etc., especial for the complicated technique like MSPC. Therefore, it is recommended to perform cost-benefit analysis while considering the nature the process and the current process performance. The idea was stimulated from our practical survey. The value of SPC or even MSPC can reveal only when its contribution can reflect on the overall benefit of the company. “Industry is not interested in the latest technological offering, be it a smart sensor or the latest distributed control system, unless its money-making potential is clear and demonstrable” (Anderson, 1997). Implementing MSPC in the right timing with proper circumstance can achieve higher utilization. Management Involvement. Statistical process control is not simply a collection of various statistics tools. Management involvement and commitment to the quality-improvement process are necessary components of a successful implementation. The manager must have strong motivation and knowledge to continuously push the implementation, on the other hand the employees need to be re-educated and encouraged. In the long term, the culture of quality oriented should be stimulated and become part of the origination. Organization of Quality Team. To implement complex technique like MSPC, expertise is a key element. Most of the companies concern the lack of knowledgeable employees. Hierarchical organization of quality management department can be a good advice. Higher level statistician should have sufficient MSPC knowledge, whereas lower level operators should receive less theoretical but more practical knowledge. Such organization is more efficient in terms of human resource expenditure and the company is also able to perform the internal training to improve the quality of employees. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 68 Chapter 6. Conclusions and Recommendations 6.3. Future Research Prospect The scope and the depth of this research are limited due to the time constraint, however, we are getting started and making the first move. Based on the reflection and the inspiration from this research work, two potential research prospects are provided as follows. Software Development. We have been pointed out that computer software is a big boost to accomplish the MSPC implementation. It is understandable and realistic to apply complex technique with the aid of IT system. The program should be developed for MSPC diagnosis and it will be even better to integrate automatic alarm system. So the software can perform the MSPC calculation, and provide quick signal and clear information to the operator. The operator will be informed where to look at, and what the possible solutions are. Potential Benefit Analysis. Although we have conducted several interviews with companies, the data was not enough for quantitative analysis. A detailed investigation regarding the relation between the potential benefit of applying MSPC and the type of production can be good evidence to motivate the company to implement MSPC. The analysis also can be used as a guideline for company’s self-appraisal, so the company can have a clear vision and more confidence on MSPC implementation. The research is temporarily summed up due to the time limit; however, it always can be improved with additional effort in the future. Finally, I would like to thank again all the contributors who did a great support on this research. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 69 Reference Reference Alt, F.B., (1985), Multivariate quality control, Encyclopedia of the statistical sciences, Vol.6, pp.111-122. Anderson, J., (1997), Future directions of R & D, Computers in Industry 34 (pp. 161-172.) Donnell, A.J., (1996), SPC implementation for improving product quality, IEEE/CPMT International manufacturing technology symposium. pp.416-421 Does, R.J.M.M., Roes, K.C.B. & Trip, A., (1996), Statistical process control in industry – Implementation and assurance of SPC, Dordrecht, Kluwer Academic Publishers. Fuchs, C. & Kenett, R.S., (1998), Multivariate quality control, New York: Marcel Dekker, INC. Hotelling, H., (1931), The generalization of student's ratio. The Annals of Mathematical Statistics, v.2, pp.360-378. Hawkins, D.M. (1991), Multivariate quality control based on regression-adjusted variables, Technometrics, Vol.33, No.1, pp.61-75. Hashmi, K., Introduction and Implementation of Total Quality Management (TQM), http://www.isixsigma.com/library/content/c031008a.asp#about, retrieved on 12.07.2005, Copyright 2000-2005 iSixSigma LLC. Kanthanathan, M. & Wheeler, S., (1990), Implementing SPC in a large manufacturing facility: An example, IEEE/CHEM IEMT Symposium, pp. 200-203 Kourti, T. & MacGregor, J.F., (1995), Statistical process control of multivariate processes, Control Eng. Practice, Vol.3, No.3, pp.403-414. Kourti, T. & MacGregor, J.F., (1996), Multivariate SPC methods for process and product monitoring, Journal of Quality Technology, Vol.28, No.4, pp.409-427. Ledolter, J. & Burrill, C.W. (1999), Statistical quality control – Strategies and tools for continual improvement, N.Y.: John Wiley & Sons, Inc. Montgomery, D.C., (1985), Introduction to statistical quality control, N.Y.: John Wiley & Sons, Inc. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 70 Reference Montgomery, D.C. & Runger, G.C., (2003), Applied statistics and probability for engineers, N.Y.: John Wiley & Sons, Inc. Mason, R.L., Tracy, N.D. & Young, J.C., (1992), Multivariate control charts for individual observations, Journal of Quality Technology, Vol.24, No.2, pp.88-95. Mason, R.L., Tracy, N.D. & Young, J.C., (1995), Decomposition of T2 for multivariate control chart interpretation, Journal of Quality Technology, Vol.27, No.2, pp.99-108. Mason, R.L., Tracy, N.D. & Young, J.C., (1997), A practical approach for interpreting multivariate T2 control chart signals, Journal of Quality Technology, Vol.29, No.4, pp.396-406. Mason, R.L. & Young, J.C., (1999), Improving the sensitivity of the T2 statistic in multivariate process control, Journal of Quality Technology, Vol.31, No.2, pp.155-165. Mason, R.L. & Young, J.C., (2001), Implementing multivariate statistical process control using Hotelling’s T2 statistic, Quality Progress, Vol.34, No.4, pp.71-73. Mason, R.L., Chou Y., Sullivan, J.H., Stoumbos, Z.G., & Young, J.C., (2003), Systematic patterns in T2 charts, Journal of Quality Technology, Vol.35, No.1, pp.47-58. Richard, A.J. & Dean, W.W., (2002), Applied multivariate statistical analysis, N.J.: Prentice Hall. Shewhart, W.A., (1931), Economic control of quality of manufactured product, New York: Van Nostrand company. Stephen, V. & Marcus, J., (1999), Statistical quality assurance methods for engineer, New York: John Wiley & Sons. Yang, K. (2004), Multivariate statistical methods and Six-Sigma, Int. J. Six Sigma and Competitive Advantage, Vol. 1, No.1. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 71 Appendix A. Formulas of Shewhart Control Charts. Appendix A. Formulas of Shewhart Control Charts. p-chart The p-chart graphs the proportions of defective items from successive subgroups. The center line and the control limits of the control chart fro fraction nonconforming are shown as follows. p(1 − p) n UCL = p + (3) Center line = p p(1 − p) n LCL = p − (3) m p= m ∑ D ∑ p$ i =1 mn i = i =1 i m D p$ i = i n i = 1,2,...m p Is the average of the sample fractions nonconforming. p$ i is the sample fraction nonconforming, which is the ratio of the number nonconforming units in the sample D to the sample size n. Di is the number of nonconforming units of a sample. n is the sample size. m is the number of samplings. As a general rule, m should be 20 or 25 (Montgomery, 1985). np-chart The np-chat is slightly different from p-chat. Instead of plotting the proportions of defective items, the number of defectives np is plotted. UCL = np + (3) np(1 − p) Center line = np LCL = np − (3) np(1 − p) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 72 Appendix A. Formulas of Shewhart Control Charts. C-chart The c-chart is applicable when the large products are inspected. The quality can be monitored by sampling the products and count the number of defects on each product. Ci is the number of defects on each product (sample size is 1). The center line and the control limits of the c-chart are shown as follows. UCL = c + 3 c Center line = c c= (c1 + c 2 + ......c m ) m for m inspections. LCL = c − 3 c U-chart The u-chart is a modification of the c-chart. The number of nonconformities per unit (ui = ci / ni) is plotted, so the sample size does not need to be one. UCL = u + 3 u / ni m Center line = u where u = ∑u i i m for m inspections. LCL = u − 3 u / ni X-bar chart and R-chart The x-bar chart plots the sample averages of sample size n over time. A set of sample data were taken from a period of stable process, it is usually recommended that number of subgroups k is at least 20 to 25 and the sample size n would be 4, 5 or 6. (Montgomery, 1985). Suppose a quality characteristic is normally distributed with mean µ and standard deviation σ. The average of a particular sample with size n can be calculated as, x= x1 + x 2 + ... + xn n It is known that x is normally distrusted with mean µ and standard deviation σ x = the center line and the control limits of x-bar chart can be expressed as, Master Thesis: Multivariate Statistical Process Control in Industrial Plants. σ n . Thus, Page 73 Appendix A. Formulas of Shewhart Control Charts. UCL x = µ + 3σ x = µ + 3 σ n CL x = µ LCL x = µ − 3σ x = µ − 3 σ n However, the mean µ and standard deviation σ are usually unknown. The estimate process average µ can be derived from the grand average, and this will be used as center line of x-bar chart. CL = x = (x1 + x 2 + .... + xm ) m To construct the control limits, σ also needs to be estimated since it is unknown. Suppose x1, x2, … , xn is a sample of size n, then the range of this sample is the difference between largest and smallest observations. R= xmax - xmin Let R1, R2, …., Rm be the ranges of the m samples. So the averages range can be calculated as, R= (R1 + R2 + .... + Rm ) m The relation between estimated σ and the average range is computed as, $= R σ d2 Therefore under the circumstance that µ and σ are unknown, the center line and the control limits of x-bar chart can be expressed as, UCL x = x + 3 d2 n R = x + A2R CL x = x LCL x = x − 3 d2 n R = x − A2R Master Thesis: Multivariate Statistical Process Control in Industrial Plants. where A 2 = 3 d2 n Page 74 Appendix A. Formulas of Shewhart Control Charts. The constant A2 is tabulated in Appendix B for different sample size n. Since the standard deviation can be estimated by sample range, we can also construct R-chart to monitor the variability of a process. The standard deviation of R is σR=d3σ. Since σ is unknown, $ R = d R . Thus the formulas of center line and control limits of so σR will be estimated by σ 3 d2 R-chart are summarized as, $ R = R + 3d R = R D UCLR = R + 3σ 3 4 d2 where D4 = 1 + 3 d3 d2 where D3 = 1 − 3 d3 d2 CLR = R $ R = R − 3d R = R D LCLR = R − 3σ 3 3 d2 The constant D3 and D4 are tabulated in Appendix B for different sample size n. X-bar chart and S-chart The mean and the variability of a process also can be monitored by constructing x-bar chart and S-chart. Instead of using the range R, the process standard deviation is used directly to monitor the variability. Generally, x-bar chart and S-chart are preferable when either (1). the sample size n is moderately large, say n>10 or 12, (2). the sample size n is variable (Montgomery, 1985). Again, the standard deviation of the process σ needs to be estimated from historical data. The average of the m standard deviations is m S= ∑S i =1 i m Where m is the number of samplings, and Si is the standard deviation of the ith sample. The formulas of center line and control limits of S-chart are summarized as, UCL S = S + 3σs = S + 3σ 1 − c 4 2 = S + 3 S 1 − c 42 c4 CLS = S LCL S = S − 3σs = S − 3σ 1 − c 4 2 = S − 3 S 1 − c 42 c4 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. $= S where σ s = σ 1 − c 4 2 and σ c4 Page 75 Appendix A. Formulas of Shewhart Control Charts. The constants B3 and B4 are defined as, B4 = 1 − 3 1 − c 42 c4 B4 = 1 + 3 1 − c 42 c4 So the formulas of S-chart can be re-write as, UCL S = B4 S CLS = S LCL S = B3 S $= S Since the standard deviation σ is estimated as σ c4 chart can be expressed as, 3S UCL x = x + c4 n , the formulas of corresponding x-bar = x + A3 S CL x = x LCL x = x − 3S c4 n = x − A3 S where A 3 = 3 c4 n The constant B3, B4 and A3 are tabulated in Appendix B for different sample size n. I-chart In the case that the sample size n is equal to 1, which means the sample only has one individual unit. Then the control chart for individual measurement I-chart should be used. The process variability is estimated with the moving range MR=│Xi – Xi-1│. The formulas of center line and control limits of I-chart are summarized as, UCL = X + 3 MR d2 CL = X LCL = X − 3 MR d2 If a moving range of n=2 observation is used, then d2 =1.128 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 76 Appendix B. Constants for Selected Control Charts Appendix B. Constants for Selected Control Charts (source: Ledolter, J. & Burrill, C.W. (1999), Statistical quality control – Strategies and tools for continual improvement) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 77 Appendix C. Hawkins’ Data Set and T2 Statistics of Measurements. Appendix C. Hawkins’ Data Set and T2 Statistics of Measurements. No. X 1 X 2 X 3 X 4 X 5 1 17.265 11.788 15.101 13.903 10.465 2 17.384 6.996 11.552 7.253 6.641 3 16.517 10.277 11.724 13.013 9.111 4 14.997 10.682 12.087 11.457 6.320 5 17.633 9.348 12.672 10.475 5.481 6 16.041 11.320 13.957 11.474 8.176 7 15.339 10.384 12.313 9.401 7.252 8 17.144 12.254 14.931 13.715 11.135 9 20.351 10.028 14.271 11.124 8.994 10 19.586 11.083 15.019 12.126 9.441 11 20.153 13.100 16.231 13.628 8.780 12 18.044 9.699 11.807 11.655 7.513 13 17.041 9.748 13.576 9.333 7.316 14 17.671 13.223 15.937 15.119 12.129 15 16.306 9.140 13.239 10.982 8.900 16 15.977 9.904 12.822 9.910 7.190 17 18.517 11.401 16.883 13.162 12.861 18 16.591 12.875 14.542 13.787 7.931 19 17.576 10.686 13.072 11.257 5.933 20 17.225 8.943 13.033 9.088 6.176 21 19.234 11.575 15.192 11.809 11.418 22 19.379 10.421 13.095 11.898 7.881 23 16.009 7.478 10.291 7.207 3.160 24 15.944 10.086 14.438 10.652 6.916 25 16.541 8.197 12.520 9.586 9.304 26 18.325 7.004 12.773 8.136 5.326 27 17.652 9.930 13.904 9.747 6.105 28 16.615 11.221 14.151 12.629 10.601 29 14.606 8.542 11.834 9.587 5.790 30 19.074 9.550 13.044 11.688 9.450 31 22.449 10.093 16.306 11.806 11.239 32 18.401 11.856 13.608 10.832 7.709 33 18.556 12.174 14.111 11.965 9.074 34 17.727 10.740 13.100 11.012 8.429 35 19.141 10.033 13.524 10.800 8.383 36 17.554 9.132 11.563 10.554 6.593 37 19.564 10.784 14.200 11.909 10.681 38 20.985 10.191 15.129 11.301 9.087 39 20.745 10.781 14.403 9.469 7.451 40 15.395 7.794 10.602 8.826 5.933 41 16.014 7.928 11.872 7.197 6.649 42 15.220 12.697 15.201 13.891 10.849 43 19.978 11.101 13.687 11.554 8.284 44 20.886 9.578 13.724 10.914 11.075 45 16.323 8.711 12.084 8.534 5.715 46 16.120 10.997 14.497 12.918 11.921 47 17.037 8.362 13.290 10.097 9.484 48 13.065 11.625 14.923 12.589 12.446 49 16.188 9.140 13.284 10.991 9.126 50 22.047 10.824 14.796 10.872 9.264 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 78 Appendix C. Hawkins’ Data Set and T2 Statistics of Measurements. T2 Statistics of measurements 36 to 50 and the T2 plot. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 79 Appendix D. Questionnaires for Interviews. Appendix D. Questionnaires for Interviews. (1). Questionnaires for Industrial Statistician Part 1: Overview of existing production Q1. How many grades/products are your producing? Q2. What is the final product of the production mentioned in Q1? Q3. How does this process operate? (Continuous, Discrete, Batch or combination.) Q4. What are the important process parameters? How are they selected? Are they in the same step or different steps of the production? Q5. How often are these process produced (if batch-wise)? Part 2: Case study Q6. Can you please mention some examples of process monitoring in the plant? Q7. Please describe the quality management system of the production line mentioned in Q6. (For example what process monitoring technique is applied (if any), what is the frequency of sampling, what is the sample size, what are the further actions after the sample analysis?) Q8. What is the existing quality related performance of the process? (For example what is the probability of abnormal situation? How often/How “abnormal” they are? What is the probability of process conformity?) Q9. What are the control limits of the monitoring system? (For example, how many sigma?) Q10. If SPC (Statistical Process Control) is applicable, do you think these SPC techniques can contribute to the improvement of operational performance of your plant? Part 3: Correlation of monitored quality parameters Q11. Do you check the correlation among the process parameters being monitored? Why do you think they are correlated? Are these process parameters present in one step or in various steps? Q12. How significant is the correlation among the quality parameters being monitored? (correlation >0.3 or 0.6 for instance.) Q13. How do you cope with the correlation of the process parameters during the monitoring? For example, what kind of technique is applied, please describe. Q14. Do you know the wrong ratio of the correlated process parameters can be an indication of process out-of-control? Q15. Do you monitor the systematic pattern in the process behavior? How do further cope with it? Part 4: Out-of-Control-Action-Plan (OCAP) Q16. How do you react when the abnormal situation occurs? How do you identify which Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 80 Appendix D. Questionnaires for Interviews. variable or set of variables or other reason (for example anti-correlation) to be responsible for the abnormal measurement? Q17. How is the response of the process after taking action on the abnormal situation? Part 5: MSPC application Q18. In general, what should or can be improved in you process monitoring system? Or what are the requirements of the process monitoring techniques/tools that you would prefer? (e.g. simple, easy, effective, accurate….) Q19. If MSPC can be proven more effective on detection of process abnormal situation for your plant, would you like to adopt it? If not, what are the barriers and difficulties (e.g. too complex, difficult to interpret the result or else)? Q20. We are conducting a research on the MSPC implementation for the industrial plants. Would you like to cooperate with our research work? We may need your comments to verify the research output and make it more applicable for the industrial plants. (2). Questionnaires for SPC Consultants Part 1: General information Q1. How many studies in the area of SPC have been performed by your company? Q2. Do you have a standard approach to perform such studies? Q3. How long are average projects performed by you with SPC? (For example, 1 month, 6 months, or else) Q4. What are the most difficult steps during performing these projects? Q5. Why the industrial plant requested you to study or to support them on the SPC? (For example, lacking of knowledge, techniques, or else.) Part 2: Production monitoring system of a selected case Q6. Can you please mention some examples of process monitoring in the plant that your company studied before? (Applying multivariate statistical process control would be preferable) Q7. How does this process operate? (Continuous, Discrete, Batch or combination.) Q8. What is the final product of the production mentioned in Q6? Q9. Please describe the overview of the process monitoring system of the case mentioned in Q6. (For example what process monitoring technique is applied (if any), what is the frequency of sampling, what is the sample size, what are the further actions after the sample analysis?) Q10. What was the old quality related performance of the process? (For example what was the probability of abnormal situation? How often/How “abnormal” they were? What was the probability of process conformity?) Q11. What were the control limits of the monitoring system? (For example, how many sigma?) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 81 Appendix D. Questionnaires for Interviews. Part 3: Correlation of monitored process quality parameters Q12. How many process parameters were monitored? Q13. How did you consider the correlation of these variables during the monitoring? For example what kind of technique was applied, please describe. Q14. How did you cope with the pattern behavior in the process parameters? Q15. In practice, the relation between the process parameters might be non-linear. Based on Hotelling’s T2 methods which only taking linear relation (positive/negative correlation) into account, can you please comment on it? Part 4: Out-of-Control-Action-Plan (OCAP) Q16. How did you react when the abnormal situation occurred? How did you identify which variable or set of variables or other reason (for example anti-correlation) to be responsible for the abnormal situation? Q17. How was the response of the process after taking action on the abnormal situation? Part 5: MSPC application Q18. Do you know these techniques? What is your impression about the applicability in the industrial situation? What are the constraints/conditions of the applicability for these methods? Q19. Based on your information, what is the level of applicability of MSPC in industrial plant and why? (For example, it is commonly used, rarely, depends on the types of production or there are certain conditions to be met.) Q20. Being familiar with the characteristics of MSPC (especially Hotelling’s T2 method), what is your opinion of the applicability in the industrial situation? Do you see any particular types of manufacturing industry are suitable (or not suitable) for this technique? Q21. We are conducting a research on the MSPC implementation for the industrial plants. Would you like to cooperate with our research work? We may need your comments to verify the research output and make it more applicable for the industrial plants. (3). Questionnaires for SPC Statisticians Part 1: MSPC from a scientific point of view Q1. Please comment on MSPC (Multivariate statistical process control), especially on the Hotelling’s T2 method? (For example, characteristics, advantages, shortcomings…) Q2. Hotelling’s T2 method has not become widely popular in the industrial plants, since it was developed. Can you please comment on it? What are possible reasons for this situation? Q3. Hotelling’s T2 method started to be addressed since 1930s, but it has been criticized for its complexity, clumsy for real application. Can you please introduce some latest development or achievement of on this concern? Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 82 Appendix D. Questionnaires for Interviews. Q4. Continued with Q3, identifying the responsible variable or set of variables for the abnormal situation can be one of the major difficulties of Hotelling’s T2 method. Can you please introduce some latest development or achievement of on this concern? Q5. Continued with Q3, the information of process pattern may not be available anymore while using Hotelling’s T2 method. Can you please introduce some latest development or achievement of on this concern? Q6. Detecting process out-of-control is one thing. Process cannot be improved if no action is taken after abnormal situation detected. Searching for the root causes of the abnormal situation becomes crucial. What recommendations can you provide on this concern? (For example, “Quality 7 tools”, or other possible techniques.) Q7. Being familiar with the characteristics of Hotelling’s T2 method), what is your opinion on the applicability in industrial plants? Do you see particular types of manufacturing industry are suitable (or not suitable) for this technique? It would be highly appreciated if you can mention some examples that you have studied. Please continue part 2 to part 5!! Part 2: Existing production monitoring system Q1. Can you please mention some examples of process monitoring in the plant that you studied before? (Applying multivariate statistical process control would be preferable) Q2. How does the product operate? (Continuous, Discrete, Batch or combination.) Q3. What is the final product of the production mentioned in Q1? Q4. Please describe the overview of the process monitoring system in the case mentioned in Q1. (For example what process monitoring technique is applied (if any), what is the frequency of sampling, what is the sample size, what are the further actions after the sample analysis?) Q5. What are the control limits of the monitoring system? (For example, how many sigma?) Q6. Why the company requested you to study or to support them on the SPC? (For example, lacking of knowledge, techniques, or else.) Part 3: Correlation of monitored quality parameters Q7. How many process parameters were selected? Q8. How did you consider the correlation during the monitoring? For example what kind of technique was applied, please describe. Q9. How did you cope with the pattern behavior in the process parameters? Part 4: Part 3: Out-of-Control-Action-Plan (OCAP) Q10. How did you react when the abnormal situation occurs? How did you identify which variable or set of variables or other reason (for example anti-correlation) to be responsible for the abnormal situation? Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 83 Appendix D. Questionnaires for Interviews. Q11. How was the response of the process after taking action on the abnormal situation? Part 5: MSPC application Q12. Do you know these techniques? What is your impression about the applicability in the industrial situation? What are the constraints/conditions of the applicability for these methods? Q13. Based on your information, what is the level of applicability of MSPC in industrial plant and why? (For example, it is commonly used, rarely, depends on the types of production or there are certain conditions to be met.) Q14. Being familiar with the characteristics of MSPC (especially Hotelling’s T2 method), what is your opinion of the applicability in the industrial situation? Do you see particular types of manufacturing industry are suitable (or not suitable) for this technique? Q15. We are conducting a research on the MSPC implementation for the industrial plants. Would you like to cooperate with our research work? We may need your comments to verify the research output and make it more applicable for the industrial plants. Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 84 Appendix E. Multiple Regression Analysis of Sub-process. Appendix E. Multiple Regression Analysis of Sub-process. Data set for sub-process analysis ORIGINAL DATA Hbtotal STANDARDIZED DATA Hbash Ret BW Hbtotal Hbash Ret BW 1 14.310 3.157 43.600 155.087 1 -1.22 -0.77 0.07 -0.47 2 14.244 3.142 43.551 154.835 2 -1.69 -1.15 0.01 -0.81 3 14.385 3.150 42.503 155.033 3 -0.70 -0.95 -1.21 -0.54 4 14.497 3.184 43.003 154.883 4 0.08 -0.08 -0.63 -0.75 5 14.297 3.157 42.978 154.037 5 -1.31 -0.77 -0.66 -1.91 6 14.517 3.172 42.914 154.748 6 0.22 -0.38 -0.73 -0.93 7 14.665 3.186 43.012 156.707 7 1.26 -0.03 -0.62 1.76 8 14.696 3.199 43.012 155.692 8 1.48 0.31 -0.62 0.36 9 14.581 3.207 43.005 154.901 9 0.67 0.51 -0.63 -0.72 10 14.548 3.205 43.016 155.217 10 0.44 0.46 -0.61 -0.29 11 14.596 3.204 42.934 154.853 11 0.78 0.44 -0.71 -0.79 12 14.490 3.197 43.018 155.422 12 0.03 0.26 -0.61 -0.01 13 14.594 3.182 43.012 154.882 13 0.76 -0.13 -0.62 -0.75 14 14.528 3.167 42.971 154.799 14 0.30 -0.51 -0.67 -0.86 15 14.516 3.165 42.985 154.938 15 0.22 -0.56 -0.65 -0.67 16 14.552 3.162 42.898 155.052 16 0.47 -0.64 -0.75 -0.51 17 14.401 3.160 42.961 154.800 17 -0.59 -0.69 -0.68 -0.86 18 14.501 3.150 43.052 154.793 18 0.11 -0.95 -0.57 -0.87 19 14.570 3.140 43.022 155.025 19 0.59 -1.21 -0.61 -0.55 20 14.452 3.132 43.037 154.800 20 -0.23 -1.41 -0.59 -0.86 21 14.483 3.165 42.939 155.315 21 -0.01 -0.56 -0.70 -0.15 22 14.513 3.180 43.076 155.030 22 0.20 -0.18 -0.54 -0.54 23 14.299 3.188 43.031 154.950 23 -1.30 0.03 -0.60 -0.65 24 14.399 3.197 43.013 154.618 24 -0.60 0.26 -0.62 -1.11 25 14.508 3.181 42.946 155.017 25 0.16 -0.15 -0.70 -0.56 26 14.420 3.157 43.012 154.800 26 -0.45 -0.77 -0.62 -0.86 27 14.552 3.140 43.071 155.321 27 0.47 -1.21 -0.55 -0.15 28 14.480 3.124 43.012 154.737 28 -0.03 -1.62 -0.62 -0.95 29 14.087 3.113 43.002 155.080 29 -2.78 -1.90 -0.63 -0.48 30 14.150 3.132 42.984 154.804 30 -2.34 -1.41 -0.65 -0.85 31 14.161 3.151 42.936 155.127 31 -2.27 -0.92 -0.71 -0.41 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 85 Appendix E. Multiple Regression Analysis of Sub-process. 32 14.528 3.196 42.899 155.380 32 0.30 0.23 -0.75 -0.06 33 14.580 3.190 42.956 155.056 33 0.66 0.08 -0.68 -0.51 34 14.560 3.182 42.926 155.453 34 0.52 -0.13 -0.72 0.04 35 14.413 3.174 43.000 154.825 35 -0.50 -0.33 -0.63 -0.83 36 14.568 3.212 43.615 157.083 36 0.58 0.64 0.09 2.27 37 14.601 3.227 44.954 157.125 37 0.81 1.03 1.65 2.33 38 14.668 3.209 45.497 156.933 38 1.28 0.56 2.28 2.07 39 14.366 3.190 45.570 156.863 39 -0.83 0.08 2.37 1.97 40 14.378 3.170 45.570 156.527 40 -0.75 -0.44 2.37 1.51 41 14.314 3.191 45.041 156.387 41 -1.20 0.10 1.75 1.32 42 14.478 3.226 44.488 156.087 42 -0.05 1.00 1.11 0.91 43 14.367 3.227 44.483 156.091 43 -0.83 1.03 1.10 0.91 44 14.503 3.266 44.520 155.901 44 0.13 2.03 1.14 0.65 45 14.690 3.253 44.472 156.177 45 1.43 1.69 1.09 1.03 46 14.752 3.275 44.569 155.923 46 1.87 2.26 1.20 0.68 47 14.583 3.272 44.511 155.804 47 0.69 2.18 1.13 0.52 48 14.585 3.255 44.574 155.934 48 0.70 1.74 1.21 0.70 49 14.646 3.239 44.444 155.949 49 1.13 1.33 1.05 0.72 50 14.580 3.227 44.565 155.964 50 0.66 1.03 1.20 0.74 51 14.573 3.233 44.485 155.997 51 0.62 1.18 1.10 0.78 Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 86 Appendix E. Multiple Regression Analysis of Sub-process. Quadratic regression model (scaled data & original data) Linear regression model with interaction term (scaled data & original data) Master Thesis: Multivariate Statistical Process Control in Industrial Plants. Page 87