200 likes | 213 Views
Process Analysis with r. BMW Group. Julia Fumbarev | 11 September 2019. Agenda. About me Process mining and bupaR Use case Next steps. About me. Business Administration and Electrical Engineering at TUM in Munich Data scientist at BMW
E N D
Process Analysis with r BMW Group Julia Fumbarev | 11 September 2019
Agenda • Aboutme • ProcessminingandbupaR • Usecase • Next steps
Aboutme • Business Administration andElectrical Engineering at TUM in Munich • Data scientist at BMW • Expertise: Machine Learning, Data Mining
Processmining • Companies collectdataabouttheirprocesses, howeveroftennotransparency • Process Mining is analyzing operational process executions based on event logs • Process mining aims at: • Discovering process model from event data • Checking the conformance of a advised model and real-life execution • Real-time process analysis • Predictive process monitoring
Processdiscovery Processmining Eventlog Processmodel * Most frequentactivities covering 60% of the event log • Algorithms: • Alpha miner • Heuristicminer • Inductiveminer • Others event = case + activity + timestamp
Tools • Commercial and open-source • Majorityare stand aloneprograms, without an interfaceto general-purposedatamining, visualizationorstatisticaltools • bupaRframeworkforeventdata (author Gert Janssenwillen, www.bupar.net) • Containspackages like edeadR (Calculatingdescriptiveprocessmetrics), processmapR (drawprocessmapandotherprocessspecificvisualization) • Easily extensible andcombinablewithothertools • Reusableanalysisscript
Usecase: Vehicledistribution Background • BMW Group distributes 2.7 Mio cars in a year; average logistic costs 5-10 €/day/car • Complexity through big number of nodes and links in the logistics network • Complexity through many external factors and manual processes Task: Identify useful patterns for deriving recommendations for optimized outplacement
Bupar - preprocessing Event filters • Activities labels • Activity frequency • Resource labels • Resource frequency • Trim cases • Trim to time window Case filters • Throughput time • Processing time • Trace length • Activity presence • End points • Precedence • Trace Frequency • Time period
Bupar - preprocessing Event filters • Activities labels • Activity frequency • Resource labels • Resource frequency • Trim cases • Trim to time window Case filters • Throughput time • Processing time • Trace length • Activity presence • End points • Precedence • Trace Frequency • Time period How do I getcarsthatstartwith a certainactivity? How do I gettraceswith 60% mostfrequentactivities?
Bupar - preprocessing Event filters • Activities labels • Activity frequency • Resource labels • Resource frequency • Trim cases • Trim to time window Case filters • Throughput time • Processing time • Trace length • Activity presence • End points • Precedence • Trace Frequency • Time period How do I getcasesfor a certainperiod? How do I getcasesforwithcertainactivities?
Usecase – controlflow • Metrics • Entry and exit points • Length of cases • Presence of activities • Rework • Visuals • Processmap • Trace explorer • Precedencematrix • Bar plot
Usecase – controlflow • Metrics • Entry and exit points • Length of cases • Presence of activities • Rework • Visuals • Processmap • Trace explorer • Precedencematrix • Bar plot Rework points to in inefficiencies (time, recourses) and shouldbeavoided Don‘tmove cars that need extra work
Usecase – Performance • Metrics • Throughput time • Processing time • Idle time • Visuals • Performance processmap • Dottedchart • Box plot Cars that were taken out of storage should have long standing time Reduce short term parking in MUCGA
Usecase – Performance • Metrics • Throughput time • Processing time • Idle time • Visuals • Performance processmap • Dottedchart • Box plot Analysis of standing and transport time
Whatelsecanbedonewiththeprocessgraph? • Clustering • Trace clustering • Sub processesclustering • Predictiveprocessanalytics • Compliance monitoring
Clustering on processgraph Trace clustering • A lot of diversity leads to complex models that are difficult to interpret • event log is split into homogeneous subsets and for each subset a process model is created • One-hot encoding of the activities and enhancing with performance metrics • PCA for reducing the number of components • K-means for splitting the traces into groups
Predictiveprocessanalytics • predicting the next activity (anditstimestamp) • predicting the future path • predicting the remaining cycle time • predicting deadline violations Applied LSTM, first results are promising. “Predictive Business Process Monitoring with LSTM Neural Networks”, NiekTax , Ilya Verenich , Marcello La Rosa , Marlon Dumas