Reliability Parametric Z-score Monitoring HanCheng Ong Design Reliability, Seagate Technology International, Singapore Reliability engineering has very much been centered on looking at the percentage of units that survives at some specified time. In fact, monitoring the parametric performance of the product is also of great importance in areas where failure rate is low or testing is expensive. By tracking the Reliability parametric, we can provide feedback to the team on the areas of weaknesses so that product improvements can be make at early design stage. A system has been build to enable Reliability parametric data to be extracted and stored in Oracle database. For each of the parameters that are being monitored, its descriptive statistics can be tabulated and charted by various product configurations. JMP scripting is used as the platform to accomplish this. The parameters can be compared by lot configurations or performance with time. In addition to the usual descriptive statistics such as mean and standard deviation, Design for Six Sigma (DFSS) measure is employed. The data set for each test configurations are distribution fitted using all available parametric models in JMP. The best-fitted distribution is used to obtain the PPM with the specification limits for computing the Z-score (DFSS measure of quality) for the parameter. The Z-scores are tabulated and charted for monitoring as indicators of quality. Results Random Seek Read Time Summary Table Zscore Mean(Rnd Seek) Table Std Dev(Rnd Seek) Table Table Run1 Run2 Run3 Run1 Run2 Run3 Run1 Run2 Run3 25C 3.6 6.4 7.8 11.1 10.7 10.6 0.3 0.2 0.2 0C 2.8 4.9 5.8 11.6 10.9 10.8 0.2 0.2 0.2 60C 2.9 6.4 6.8 11.2 10.8 10.7 0.5 0.2 0.2 Temperature Figure 1. Example of Parameter Trend Monitoring Table Random Seek Read Time Zscore Chart by Run # 9.0 8.0 7.0 6.0 Zscore Introduction 5.0 4.0 3.0 2.0 Best Fitted Parametric Model The data set for each test configurations are distribution fitted using all available parametric models in JMP. JMP provides a relatively large list of common and useful distribution models (Johnson Sl, Johnson Su, Normal, Glog, LogNormal, Exponential, Gamma, Weibull and Extreme Value). Each distribution is tested for Goodness-of-Fit with the data set. The best fitted distribution is used to obtain the PPM (parts per million) with capability analysis. The specification limits are specified using the “Set Property” function in JMP. The PPM from JMP output is then used to compute the Z-score (DFSS measure of quality) of the parameter for each configuration of interest. Non-Parametric Smoothing Model In the event that none of the available distributions shows a good fit (determined by the p-value output), the non-parametric smoothing available in JMP will be initiated for the Z-score determination. The Z-scores are tabulated and charted for monitoring as indicators of quality. JMP Syntax: Parametric Method: Column(dt, “Parameter")<<Set Property( "Spec Limits", {LSL( lsl_para )} ); Distribution( Continuous Distribution( Column( :Parameter ), Fit Distribution( All ) ), By( :Config )); Non-Parametric Smoothing Method: Distribution( Continuous Distribution( Column( :Parameter ), Fit Distribution( smooth Curve( Spec Limits( LSL(lsl_para) )))), By( :Config )); Figures 1 and 2 depicts the Trend table and charting of the Random Seek Read Time parameter at various test temperature for illustration (data are simulated). 25C 0C Run3 Run2 Run1 Run3 Run2 Run1 Run3 Run2 0.0 Run1 Methods 1.0 60C Figure 2. Example of Parameter Z-score Trend Monitoring Chart Discussion •Besides the usual statistics like mean and standard deviation, the DFSS measure of quality, Z-score is used as the monitoring metric. Z-score takes into account both the central tendency and dispersion. •The parameter tables and charts are used for monitoring and feedback to the team on the product performance. For areas where the Z-score is deemed to be low, improvement actions will be put in place and the parameter will be remeasured after the fix is implemented. •The Z-score is computed based on best fitted distribution and not normal distribution which many have erroneously assumed (whether it is for convenient sake or due to lack of knowledge). •This method of monitoring the parameter performance using Z-score is suitable during product design stage. When the product is stable and is in production mode, control chart type of monitoring may be more suitable. •Parametric Monitoring provides a good additional eye for product reliability monitoring especially where large sample testing is not viable due to good reliability and failures are hard to come by. It is also useful where it is expensive to have large samples and long test duration to detect failures. Reference JMP User, Stats and Graph, Scripting Guides