Predictive Maintenance: From POC to Production With Spark
summary
Proof of concept runs in 10 weeks (2 for discovery, 8 to build) and moves to a six‑month pilot; models use random forests and experiments with neural networks with Theano, then get ported into a Spark pipeline. Failure events are very rare, labels are noisy and timing is imprecise, and onboard systems emit ~2500 diagnostic codes, forcing heavy cleaning and feature aggregation. Random-forest models beat existing rule filters in trials (example: detecting ~6/10 vs 2/10 failures among top alerts), but neural nets prove unstable in changing production data.
AI-written summary · about
chapters (12)
- 0:00Intro
- 2:47Data Collection
- 5:00In Production
- 10:40Big Data Analysis
- 15:40Data Quality
- 20:07Train Dreams
- 26:06Data Science
- 30:24Big Data
- 31:26Prediction
- 32:36Reporting
- 33:36Real Life Questions
- 38:36Accuracy
related talks
40mPredictive Maintenance
26mProactive Failure Prediction Through Anomaly Detection
41mPredictive Maintenance with Azure Synapse
14mTransforming Manufacturing: AI-Driven Predictive Maintenance
37mForecasting using machine learning methods
42mPredicting Weather-Caused Rare Events: A Utility Outage Prediction Use Case
18mPredictive Network: How AI Forecasts System Failures
39mMachine Learning with Scala on Spark