Robert Morris, Ph.D. Co- Founder & Chief Science Officer Predikto, Inc. Deploying Predictive Analytics Solutions in the Rail Industry and Seeing a Return on Investment
SHOWING BUSINESS VALUE FROM DEPLOYED PREDICTIVE ANATYICS IS INCREDIBLY CHALLENING!
INDUSTRIAL BUSINESSES HAVE BIG DATA BUT STRUGGLE TO SEE VALUE FROM IT. WHY?
SOME CHALLENGES 1) IDENTIFYING THE LOWEST HANGING FRUIT 2) MAXIMIZE THE USE OF DESPERATE DATA SOURCES 3) NEED FOR ADAPTATION TO TEMPORAL AND SPATIAL CONTEXT CHANGES ON THE FLY 4) SCALE ACROSS A TREMENDOUS VOLUME OF USE- CASES AND BEING ACCURATE 5) SHOWING THE VALUE TO BUSINESS 6) COMMUNICATING THE VALUE OF RESULTS
BUSINESSES TEND NOT TO KNOW WHAT THEY WANT OR CAN USE PREDICT FUTURE EVENTS? FORECAST RELIABILITY? INSIGHT TO OEPRATIONAL DATA? TRIAGED MAINTENACE? IT S NOT JUST AN ANALYSIS IT S A SOLUTION
ACCURACY VS. UTILITY WERE YOU RIGHT OR WRONG? HOW RIGHT IS RIGHT? HOW ACCURATE MUST WE BE TO SHOW VALUE?
HOW DOES PREDIKTO SCALE? AUTO- DYNAMIC PREDICTIVE ANALYTICS SOFTWARE GETTING TO VALUE WITH THE DATA AVAILABLE NOW, NOT TOMORROW
HOW PREDIKTO SOFTWARE DELIVERS ACTIONABLE RESULTS EMD INTELLITRAIN GE RM&D NYAB LEADER MOTIVEPOWER WABTEC PREDIKTO ENTERPRISE PLATFORM SENSORS SAP INFOR ORACLE MAXIMO EAM PREDIKTO INPUT API S AUTO- DYNAMIC DATA TRANSFORM. ENGINE AUTO- DYNAMIC MACHINE LEARNING ENGINE PREDIKTO OUTPUT APIS & BI VIZ. ACTIONABLE PREDICTIONS WILD BEACONS WEATHER CUSTOM APPS UMLER TCIS OTHER 8
ARCHITECTURE DIAGRAM Data Aggregation/ETL Machine Learning/Analytics Outbound APIs/Integration Standard JSON Data Cleaning Email Predikto data Pipeline AutoDynamic Feature Engineering AutoDynamic Feature Selection MAX Dimension Reduction AutoDynamic Machine Learning Ensemble Operational Integration SMS WebHook ETL Post-Processing and Calibration UI Data Store Apache Spark 9
10 SAMPLE USE CASES
REDUCING BAD STOPS IN FREIGHT RAILROAD BUSINESS CHALLENGE Class 1 railroad experiencing unscheduled train stops due to false warnings from wheel monitoring sensors (HBDs) 800 HBDs installed throughout 22,000 miles of track Unscheduled railroad stops due to HBD failures costing over $10M annually RESULTS Solution entered production June 2014 12.7% reduction in the number of bad stops $1.5M impact in Year 1 Predikto identifying 37% of all bad stops with 7 day advanced notice SOLUTION Use the Predikto platform to improve reliability, reduce train delays and reduce maintenance costs Automatically predict which HBDs are expected to fail within the next 7 days. Allows maintenance teams to be proactive and focus on the right equipment and tools Dashboard with GIS maps showing bad actors Asset health score to show worst performing HBDs 11
PREDICTING FAILURES IN COMMUTER TRAIN DOORS BUSINESS CHALLENGE Malfunctioning doors on commuter trains is a costly problem for manufacturers, service providers, and passengers Manufacturers have Service License Agreements (SLAs) that result in fines for service delays due to failed equipment When a door on a commuter train does not open properly, passengers take longer to exit and board, resulting in delays, increased safety incidents, and potential for loss of revenue RESULTS Predictions of whether and when a specific door in a door set on a specific car on a specific train will fail, providing for a 7 day window. The precision (accuracy) of the warnings produced by Predikto is in excess of 84 percent. The solution reduces delays caused by door failures, enhances train travel velocity, expedites repair time, and most importantly, saves time and money. SOLUTION Actionable and tailored prediction of whether a specific train door will malfunction at some point in the future and allowed for enough time so that an actionable response to the problem could happen, but before delays occurred Malfunctions over the next 7 days delivered via the Cloud The output from the Predikto s automated solution were to be interfaced with the customer s existing data management to take action on immediately 12
PREDICTING FAILURES IN RAILROAD TRACK BUSINESS CHALLENGE Cracks or breakages in segments of track significantly increases the likelihood for derailment. Need for better methodology to prioritize track inspection. Narrowing inspection to a small track segment. SOLUTION Analysis of historical track data, usage and weather to identify risk track segments Provide granular predictions to quarter mile segments. Provide a lead-time to allow the customer to properly schedule and procure for predicted failures, thus reducing operational costs. Algorithms focused on accurate detection. RESULTS Actionable predictions on which half-mile segments of track are most likely to experience a defect well in advance of it actually occurring at some point during the following 3 months with a 63% accuracy. Predictions on defects were based on data provided by the customer including: ultrasonic readings, historical defects and track details, Class I Railroad company (train movements, tonnage, etc.), and data built in by Predikto (e.g., weather). 13
PREDICTING DELAYS IN BULLET TRAINS BUSINESS CHALLENGE European railroad experiencing stops/delays due to mechanical failure Delays greater than 20 minutes including breakdowns Cost to pull engine out of service very high TARGET RESULTS 100% Precision in Identifying 10% of Control Unit Delays Identification of failures and delays greater than 20 minutes 1-4 days in advance. Identification of level 1 and level 2 systems SOLUTION Predikto providing alerts at train, car, component, and sub-component levels Provide top most likely error codes associated with failure Algorithms configured for low false positive rate Daily notification to operations team 14
PREDICT PREVENT PERFORM