If there’s contrast enough NOVEMBER, 2018 You can directly write your title over the image. AITalk Berlin Use cases from the field Borealis at a glance 2nd largest polyolefin producer in Europe Borouge operates world’s largest integrated PO site in Ruwais, UAE Head office in Vienna, Austria 6,600 employees EUR 7.5 billion sales revenue in 2017 (>EUR 9 billion with Borouge) +3,500 in Borouge Operates in over 120 countries on 5 continents Ownership structure EUR 1,095 million Production and sales of polyolefins, base chemicals and fertilizers net profit for 2017 Mubadala, UAE OMV, Austria Agenda 01 Selection a viable use case 02 Predictive maintenance 03 Energy Optimisation 04 Key Takeaways Work package objectives Selecting a viable use case Getting ready for the future Generic approach Equipment selection decision tree Structure to select maintenance approach Initiatives to enable failure prediction on a broader scale Stakeholder interviews Impact assessment & initiatives definition Pilot & Lessons Learned User journeys Detailed initiatives Project charters Select & execute selected initiatives Enablers for failure prediction Strong contribution to Reliability and Maintenance Excellence Failure prediction High quality data Equipment selection Continuous improvement Alarms and preventive action Work processes Maintenance data input (in the field) Consistent root cause analysis Data strategy for critical equipment Alarm database Data middleware Decision tree Implement predictive model Data & tools Wireless data network Asset Portal Analytics tools (pattern recognition) Organisation & competences Communication Data fluency training Visualise model results PdM visualisation (monitoring, POT , KPI) Enablers – Organisation • Data driven organisation • • • Communication on predictive analytics Mind-set to use data to support decisions Root cause analysis verified with data (failure modes) • Resources and Competences • • User/role specific training Data analysis and statistics (basic – advanced) Reliability / Process / TD&E engineers Right analytics tools e.g. pattern recognition tools Enablers – Work Processes • Reliability Management work process • • • • Equipment strategies (min. standards/FMEA/maint.tasks) Equipment monitoring requirements (sensors, condition monitoring) Integrate also into design requirements of new investment projects Structure to select maintenance approach (decision tree): Key drivers for equipment selection • Number of failures • Data quality & equipment knowledge • Equipment criticality • Data set prediction power • Business impact Maintenance approach • Corrective maintenance • Preventive maintenance • Predictive maintenance • Unsupervised anomaly detection • Supervised anomaly detection • Maintenance Execution Excellence work process • • Structured maintenance feedback Efficient and integrated documentation (e.g. work in field) Enablers – Data • Wireless technology • • • Wireless sensors and condition monitoring Connectivity in the plants Supporting also work processes • Centralised data management • • • Mapping of Visualisation and Asset portal requirements Data collection from different sources (data middleware) Learning and usage across plants/locations Hyper Compressor – Data assessment Dimension Question asked Assessment example Score Assessment result Completeness Are all data sets and data items recorded? What is the percentage of missing values? 0.977 Ok Does the data reflect the data set? How many outliers are there? What is the percentage of negative values in positive defined variables? How many out-of-range variables are there? 0.968 Ok Accuracy Data quality examples Sensors VI52523X VI52524Y VI52525X VI52526Y PDI52034B PDI52034C PDI52034A TI52563 PI52005A PI52005B Number of missing Frequency of missing values values 319823 319824 319825 319826 205764 205765 194570 51494 41213 41203 Completeness 0.783 0.783 0.783 0.783 0.504 0.504 0.476 0.126 0.101 0.101 Sensor Number of negatives Negative score Number of outliers Outlier score Global score FI52524_1F 360778 0.117 8477 0.979 0.548 FI52524_1A 305056 0.253 33714 0.917 0.585 FI52524_1E 309738 0.242 222 0.999 0.621 FI52524_2A 214679 0.475 47977 0.883 0.679 FI52524_2C TI52572 TI52557 149868 0 0 0.633 1 1 39444 163507 159408 0.903 0.6 0.61 0.768 0.8 0.805 TIC52043_PV 0 1 137106 0.627 0.813 FIC52303_PV 13 1 125544 0.693 0.846 VI525211D 0 1 106266 0.739 0.869 VI525212D 0 1 105082 0.741 0.871 Accuracy Treatments See details on next pages Comparison Data quality remained similar to what was found in the previous assessment focused on data from 2016 Advanced Analytics – Data Feasibility Results Work package objectives Predictive Maintenance on a hypercompressor Predictive maintenance – objectives Increased plant reliability Use data to work in “earlier part of the P-F curve” Failure prediction to prevent unplanned stops Intervene pro-actively • Use data to understand the underlying root causes Continuous improvement steered by POT • Estimate remaining useful life of equipment / asset (component) • Improved process efficiency • • Data fluent employees Aligned work processes to reduce variability Process Overview LP Recycle HP Recycle B&P Compressor Hyper Compressor Reactor Separation Hyper Compressor 2 x 6 Cylinders Plunger HP Packing Central Valve Work process Connecting the physical and digital worlds to create a smarter business Continuous and automated readings by sensors Data is aggregated and processed Pilot progress Physical Assets Analytics reveal trends and patterns Processes Actions are taken to drive value Insights allow better decisions and predictions Digital Data science process Data feasibility assessment Equipment selection Data cleansing Information selection Tags selection Booster Primary (BP) Master Data Set Hyper compressor (HC) machine process total tags Model assessment Validation Failure recognition Failure classification Model building Refine model Improve model accuracy 383 730 1113 Failure/Non-failure time-window identification KNN Logistic regression Random Forest total tags 772 Implement model Going online Visualise model outcomes Dashboard development Advanced Analytics – Overview Hyper Compressor – Normal Operation Random Forrest – Hyper Compressor – CV Failure Outcome & Learnings Outcomes During normal machine operation no noticeable difference between Homo- and Heteromere production Good models detect signal of failures up to 6 days before a failure Component-specific models show strong accuracy to predict a failure 14 days in advance Top 5 learnings Most likely suitable for large & complex problems Multidisciplinary expertise required Good & consistent data quality is key High development cost, low running cost Scalability depends on modelling approach & equipment strategy Deployment: online model scoring Modelling Visualisation Daily Top 10 anomalous tags WS anomaly Quality checks KS anomaly Data Gateway Hourly Data pipeline Site IP21 Plant IP21 Quality checks SQL Server DCS kNN Feature generation Feature generation Logistic Regression Random Forest Failure probability Work package objectives Energy optimisation in a chemical plant Energy optimisation objectives To increase the energy efficiency of the Kallo Dehy plant, while maximising the overall economic value of the plant • Linking Energy KPI variability to operations decisions • Development of dynamic energy target models Energy streams to be considered are, in order of priority: 1. Natural Gas, Fuel Gas, Electricity 2. Steam 3. Compressed Air, Cooling water, Nitrogen 24 | Approach ENERGYmaestro Process & business understanding Economic mapping & financial modelling Workshops with operators Steam extraction down Capacity of steam extraction Advanced analytics Implementation preparation & feedback Training & review Maintenance – waiting for spare parts SO2 alarm Sulfine plant is not in operation Lack of sulfur There is no steam for the turbine Low pressure boilers at P2 Boiler BERI Steam extraction < maximum HP vers LP Capacity of LP steam network Boiler SO2 Consumers Demand/losses Start-up valves Training Overall management Communication & coordination 2 - 4 weeks 2 - 4 weeks 2 - 4 weeks 4 - 8 weeks 2 - 4 weeks 2 - 4 weeks Process & business understanding Economic mapping & financial modelling Workshops with operators Implement. preparation & feedback Advanced analytics Training & review Process understanding • • Site visit and process description Management Energy Monthly MWh/T C3= Evaluation of P&IDs / PFDs • Evaluation of data availability • Data collection Business understanding Plant Electricity Gas Steam LTRS Weekly MWh/T C3= MWh/T C3= MWh/T C3= C3= Losses Production Electricity Gas Steam LTRS Daily MWh/T C3= MWh/T C3= MWh/T C3= C3= Losses Operations Air Compressor Shift MWh/T C3 Operations • • Identify key financial data and understanding Data collection Real-Time Key process variables Splitter Compressor MWh/T C3= Key process variables Charge Heater Air Heater Nuovo Pignone LTRS MWh/T C3 MWh/T C3= MWh/T C3 C3= Losses Key process Variables Key process variables Key process variables Key process variables Key process variables to be determined by advanced analytics Economic mapping & financial modelling Process & business understanding • Workshops with operators Advanced analytics Implement. preparation & feedback Analyse and quantify historical variability against baseline target of initial KPIs. Training & review Process & business understanding Economic mapping & financial modelling Workshops with operators Implement. preparation & feedback Advanced analytics Training & review • Workshops with operators and plant staff • Brainstorming root causes of variability and process improvement ideas • Objective to engage operators in the project • Ideas presented on a cause tree • Cause tree communicated to operators Steam extraction down Capacity of steam extraction Maintenance – waiting for spare parts SO2 alarm Sulfine plant is not in operation Lack of sulfur There is no steam for the turbine Low pressure boilers at P2 Boiler BERI Steam extraction < maximum HP vers LP Capacity of LP steam network Boiler SO2 Consumers Demand/losses Start-up valves Training Overall management Communication & coordination Economic mapping & financial modelling Workshops with operators Advanced analytics Implement. preparation & feedback Collaboration with plant staff Pareto analyses to identify top influencing parameters Training & review Decision trees to map impact of top parameters on efficiency Factor Yield curve to show change in efficiency versus multiple variables Top Explanatory Variables: ▪ Production rate ▪ Maintenance operations ▪ Impact of input feed quality ▪ Product quality giveaway ▪ Reactor temperature ▪ Steam extraction ▪ Input pressure Production rate (T/Hr) Histograms to diagnose statistical splits in performance Number of hours Process & business understanding Composition (%) Dynamic Target Advanced analytics: root causes • Decision Tree based on good and poor performance periods • Discussion with process experts to determine controllable variables Decision Tree for LTRS Ratio (Greater or less than 1.1% Propylene Losses) 30 | Workshops with operators PD F- X C h a n Implement. preparation & feedback Advanced analytics Training & review ge w PD F- X C h a n ge ww w W! y NO bu to ck Cli .d o oc u-tr a c k.c m ww ! NOW buy kto Clic .d o oc u -t r a c k.c m Economic mapping & financial modelling Process & business understanding Engis D1 - Causes de surconsomma on vapeur Surconsomma on vapeur (EUR) 6000 5000 4000 Oxyda on 3000 Extrac on D1 Défluora on 2000 1000 0 Arrêt By-pass Comme ntaires : ............. : ............. ............. ............. ............. ............. ............. ............. ............. ............. ............. ............. ............. ............. ...: ............. ............. ............. ............. ... ...: • Develop useful indicators in control screens • Develop daily/weekly/monthly reporting • Develop energy implementation roadmap • Feedback to operators, plant staff, engineers and managers Capacité Process & business understanding Economic mapping & financial modelling Workshops with operators Advanced analytics Implement. preparation & feedback Training & review Work package objectives Key Takeaways Key Takeaways • Business understanding is key • • • • Apply subject matter expertise in data preparation Reduce number of variables based on SME Create meaningful features for the problem at hand Correct the target you want to predict based on SME • Reduce complexity: • start with a classifier iso a regressor • Use explainable models • Run in iterations (agile approach) • Always predict a target that can be converted to actions