SAP Predictive Maintenance and Service, n-premise editin and Data Science Custmer
Predictive Maintenance is a prcess, nt just an algrithm Dmain expertise is as imprtant as data science, if nt mre * * * * * * CRISP-DM Crss Industry Standard Prcess fr Data Mining * Invlvement f dmain expert necessary 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 2
custm custm custm Data Science Predictive Maintenance Architecture Actins Business Prcesses Business user Insights PdMS Applicatins (Insight Prvider mash-up, use case-specific) Insight Prviders (peratinalized data science, micr-services) Dmain Expert Derived Signals Data Analyst / Scientist Raw Data Data Management & Data Science (lambda architecture, schema n read) Data Manager Operatinalize Data Science by embedding it in applicatins and use it t trigger fllw-up actins. Data Science algrithms frm many surces R, PAL, APL plus custmer s wn. 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 3
Data Science Predictive Maintenance Architecture 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 4
Data Science Predictive Maintenance Architecture 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 5
Data Science Predictive Maintenance Analyses Defect Pattern Identificatin Statistical analysis, text clustering, assciatin analysis, and decisin trees. Visualizatin f big data with parallel crdinates and multi-dimensinal scaling. Maintenance Priritizatin Frecast prductin and prbability f failure calculatin at asset level. Priritize maintenance activities based n actual and frecasted KPIs Systems Trending and Alert Management Detect utliers and anmalies in the data with supervised and unsupervised machine learning Text analysis and text mining t classify scheduled vs. unscheduled maintenance events Machine Health Predictin Histric machine data are used t predict breakdwns via decisin trees Energy cnsumptin pattern prfiles are calculated with k-means clustering Dmain expert knwledge was mdelled in SAP HANA with decisin tables Vehicle Health Predictin Used assciatin rule mining and regressin tree learning t crrelate prductin rewrk and custmer satisfactin data SAP HANA/R data mining and data visualizatin capabilities applied t surveys and prductin data sets Health Predictin fr aircraft cmpnents 5 aircrafts ver 5 years, 400 sensrs each aircraft 44.1 Billin sensr readings Crrelated with maintenance histry (ntificatins), weather data & ge lcatins Text Analysis t understand maintenance activities Use f Statistical Prcess Cntrl and Symblic Aggregate Apprximatin fr anmaly detectin Bad Actr Analytics Weibull life time analyses Use classificatin techniques t identify rtating equipment likely t fail based n past patterns Rt Cause Analysis fr Quality Issues Find causal relatinships between claims and prductin settings frm machine readings Imprve n Statistical Prcess Cntrl usage Emerging Issues Analysing telematics data & relating them t equipment's service and warranty data using text mining, assciatin analysis Predictive Quality Assurance Visual detectin f cracks (image prcessing techniques) Heat image cmparisn f areas f interest f sample images f material with issues t current material (Euclidean vectr distance calculatin) Maximize Machine Efficiency in Prductin Augmenting (human) expert rules with (machine) rule mining (regressin trees) Apprximating machine state t circumvent rare event prblem (anmaly detectin) De-clutter sensr data fr rt cause analysis (trend analysis) Asset Health Predictin Optimize testing and crew efficiency based n limited resurces Optimize capital investment fr URD Cables Machine health predictin frm histric data and utages 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 6
Data Science Predictive Maintenance Algrithms Autmatic Rule Extractin with Decisin Trees Principal Cmpnent Analysis f Switches Expert Rule Validatin Weibull Remaining Useful Life Estimatin Wasserstein Metric fr battery perfrmance analysis PCA Anmaly Detectin with Principal Cmpnent Analysis scres K-Medid cluster analysis t partitin the ppulatin int classes f similar devices Battery behaviural grupings 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 7
Many Data Science Algrithms are used in Predictive Maintenance Autmatic Rule Extractin with Decisin Trees Principal Cmpnent Analysis f Switches Expert Rule Validatin Weibull Remaining Useful Life Estimatin Wasserstein Metric fr battery perfrmance analysis PCA Anmaly Detectin with Principal Cmpnent Analysis scres K-Medid cluster analysis t partitin the ppulatin int classes f similar devices Battery behaviural grupings 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 8
Data Science in Predictive Maintenance (PdMS n Premise) PdMS n Premise prvides three Data Science services ut f the bx that can be applied t custmer data. - Anmaly Detectin, Distance-Based Failure Analysis and Remaining Useful Life Predictin Anmaly Detectin Distance-Based Failure Analysis Remaining Useful Life Predictin Custm Data Science Service New Data Science Services (algrithms) can be integrated by embedding the algrithms using R packages that cnfrm t the interface prvided. PdMS n Premise includes functinality t manage Data Mining Mdels 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 9
Data Science Services Anmaly Detectin Business Prblem: Certain cmpnents are crucial fr the functining f a machine r are very expensive, therefre abnrmal behavir shuld be detected and ptential prblems fixed. Minimizing cst and dwntime are imprtant gals as well. Slutin: 1. Data extractin and preparatin: Extract the equipment s senr data frm the time series strage and prepare it. 2. Learn a mdel and stre it: Apply Principal Cmpnent Analysis t the histric sensr data f healthy machines. 3. Apply mdel t new data and stre results (anmaly scres) in the time series strage. 4. Shw anmalies in the applicatin and allw t create wrk rders. PCA Benefits: Autmatic detectin f multivariate anmalies which culd lead t failures in cmpnents and pssibility t take actin. Helps t prevent dwntimes and minimize maintenance csts. 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 10
Data Science Services Distance-Based Failure Analysis Prblem: Sme cmpnents are crucial assets that can lead t unexpected dwntimes when they fail. Slutin: 1. Data extractin and preparatin: Extract the equipment s senr data frm time series strage and prepare it. 2. Learn a mdel and scre new data: Cmpute the distances f each cmpnent t a reference cmpnent using earth mver s distance (lazy learner type f algrithm) and stre distances in time series strage. 3. Shw cmpnents ranked by distance (in descending rder) in the applicatin. discharging Bad cnditin full charging Benefits: Early identificatin f malfunctining cmpnents in rder t reduce dwntime. Reference cmpnent signature Specific cmpnent signature 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 11
Data Science Services Remaining Useful Life Predictin Prblem: The frequent ccurrence f failures leads t increased csts fr maintenance and dwntime. Slutin 1. Data extractin and preparatin: Extract repair data (IT data) and cmpute repair time KPI s (mean time between repairs, uptime, dwntime) fr cmpnents. 2. Learn a mdel and stre it: Perfrm Weibull Life Time Analysis n repair data and stre mdel. 3. Scre new data: Calculate remaining useful life and prbability f failure fr each machine/cmpnent and stre scres in time series strage. 4. Shw remaining useful life and prbability f failure in applicatin. Benefits Use statistics based apprach fr estimating lifetime f cmpnents. Usage f Weibull Life Time Analysis beneficial if n sensr data available (e.g. new cmpnent). 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 12
Predictive Maintenance Discver rules r relatinships between attributes and failures Find the reasn Incident Rt cause analysis Extract rules Required: An incident indicatin (message r failure criteria) Attributes with ptential relatin t failure (sensr data, cnfiguratins, cuntries, prduct lines, ) Usage: Serves as starting pint fr expert review Prduct/prcess imprvement based n analysis results Algrithms: Decisin Tree, Assciatin Rule Mining Permanently remve reasn if pssible cuntry A Nt A engine air cnditining 50kW 150kW yes Incident Discvery n Frequent Failures Cuntry A + Air Cnditin A A B A C A B X X X X X X A -> X 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 13
Predictive Maintenance Detect anmalies that led t upcming cmpnent failures Required: An incident indicatin (message r failure criteria) Attributes with ptential relatin t failure (sensr data, cnfiguratins, cuntries, prduct lines, ) Anmaly Detectin Autmatically find early deviatins frm nrmal Usage: Early failure indicatr Prduct/prcess imprvement and understanding Algrithms: Univariate methds (Bx Plt Rule, Statistical Cntrl Charts, Student s Test, Grub s Test, Likelihd Rati Test ) Multivariate methds (PCA, One-Class Supprt Vectr Machines, Self-Organizing Maps, Neural Netwrks) Analyze and repair if necessary 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 14
Predictive Maintenance Create statistical mdels based n past sensr data and failures Classificatin Learn a predictin Mdel frm past sensr data and failures Required: Usage: A classificatin f past data (failure / n failure) Attributes with ptential relatin t failure (sensr data, cnfiguratins, cuntries, prduct lines, ) Sufficient number f failures that ccurred in the past that can be related t the sensr data Optimize maintenance t reduce csts and dwntime Optimize spare parts planning Algrithms: Decisin Trees, Lgistic Regressin, Supprt Vectr Machines, Neural Netwrks, Binary Classificatin Repair if prbability f failure high 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 15
Example: Frm battery data t battery health t actin fr Trenitalia Cllect Real Time Telemetry Data frm Batteries Fleet Level Infrmatin Actin ERP + MRS Machine Level Infrmatin Insights Battery Real Time Data = Ideal State Maintenance Schedules & Financial Infrmatin Cmpare and analyze real time battery perfrmance t ideal state PdMS Applicatin shws insights abut health scre f machines and business infrmatin Outcme 60% maintenance cst reductin fr batteries 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 16
Data Science Predictive Maintenance Applicatin Machine Health Cntrl Center - Machine Explrer Machine Health Cntrl Center - Anmaly Scre Histry Machine Health Cntrl Center - Gespatial Visualizatin Machine Health Cntrl Center - 3D Data Visualizatin 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 17
Data Science Predictive Maintenance Applicatin Health Fact Sheet Health Fact Sheet by Cmpnents Risk Matrix and Survival Curve MRS Planning bard with re scheduled maintenance 2016 SAP SE r an SAP affiliate cmpany. All rights reserved. Custmer 18