Predictive maintenance

Overhaul

Predicts how many more flights each engine in a made-up regional airline's fleet can make before it needs to come off, and how sure it is. A deep learning model trained on NASA's turbofan run-to-failure data gives every engine a remaining life with an 80% band, and a live fleet replay lets you plan the removals: pull too early and you waste engine life, too late and an engine fails on wing and grounds the aircraft. The same approach works for any equipment that logs sensor readings until it wears out, such as pumps, generators, compressors or a delivery fleet.

Model page: test RMSE of 13.1 cycles, predicted versus actual remaining life for 200 held-out NASA engines, band calibration by remaining life, and a comparison with a single CNN and a LightGBM baseline

Removals page after a month on the auto-planner: 17 planned removals, none failed on wing, 90% of life used, and the removal log with each engine's flown and remaining cycles

What I built

A Python training pipeline on NASA C-MAPSS (200 engines flown to failure, 14 sensors, two fault modes): smoothed 50-cycle sensor windows, a LightGBM baseline, and an ensemble of five dilated 1D convolutional networks in PyTorch that each predict the 10th, 50th and 90th percentile of remaining life
Honest evaluation: settings chosen on validation engines only, then scored on NASA's 200 held-out test engines. The median is off by about 13 flights (RMSE), level with the LightGBM baseline, and the band holds the true value 78% of the time against an 80% target
The model exported to a 490 KB ONNX file and run in the visitor's browser with onnxruntime-web: about 100 ms to score 200 engines, no inference server
A "why it's flagged" view that resets one sensor at a time to how it read when the engine was installed and shows how much life the model gives back, computed live in the browser
A fleet replay of 12 aircraft and 24 engines flown one flight a minute, where every prediction comes from a model trained without that engine (5-fold, out of fold). Visitors book removals for an overnight, set an auto rule, fast-forward a month and see planned removals, failures, ground time and cost
A backtest across all 200 engines: pulling when the cautious estimate drops to 15 flights lets none fail on wing and uses 90% of each engine's life, while a fixed interval with the same failure rate uses 59%

Stack

Python, PyTorch, LightGBM, scikit-learn, pandas, ONNX, TypeScript, SQLite

Live

maintenance-demo.jnathan.dev ↗