Business Objective
Our client is a leading North American railroad company. The client’s Hot Bearing Detectors (HBD) process consists of wayside sensors detecting bearing health by monitoring its temperature and raising alerts in case of faults. The current rules of fault detection generate a lot of false alerts resulting in train delays and loss of money. There were no reports/dashboards to view/analyze the consolidated results.
The client hence wanted to build an HBD monitoring system to
• Enhance the performance of the existing talker alert system
• Build new rules to augment the current alert system
• Identify faulty detectors
Challenges
- Sparse and fragmented data, stored in multiple locations
- Due to gaps and periodicity of purge cycles, data is limited to alerting trains
- Only 80% of the trains passing have valid train ids
- Different retention rules depending on data granularity
Solution Methodology
- Input data included detector readings, alerts data, inspection data, car repair billing, detector installation data
- Enhanced current talker alert system by separating business rules for each type of detector and model with three key variables: number of cars with alerts, the average temperature difference between both sides of the train, and temperature deviation from the train average. Leveraged Decision tree technique and implemented model within the detector, on top of the existing alert system
- Built new alert rules to augment the current system and developed a single model with 20 variables, 2 stage ensemble models using XGBoost and implemented it at a central location, triggered after the existing alert
- Defined Faulty Detectors and performed descriptive analytics to identify faulty detectors
- Developed an interactive dashboard providing result classification by detector location, detector type, alert type, train stops, and time
Business Impact
- Model and manufacturer-wise modeling on existing rules led to ~20% reduction in false alerts and captured 100% true alerts
- New alert rules developed over existing rules enhanced the performance further by reducing false alerts by ~46% and capturing ~97% of true alerts
- Built a faulty detector analysis interactive dashboard which helped the client save significant operational costs because of delays and unfortunate events
- Identified data gaps and discrepancies at multiple places in existing logic documentation
