ReneWind Predictive Maintenance

Summary

A case study from the UT Austin and Great Learning PGP in AI and Machine Learning. Built a classification model to predict wind turbine generator failures using historical sensor data from 40 instruments per turbine, testing 7 model configurations.

Accuracy is the wrong headline for this problem. Failures are rare, so a model that predicted “no failure” every time would still score in the high 90s while being useless. The metrics that matter are recall, 85% of true failures caught, and precision, 97% of flagged failures were real. The model was tuned deliberately toward recall, because a missed failure costs a generator replacement while a false alarm costs an inspection.

Approach

Trained neural networks across 7 configurations, varying depth, optimizer, dropout and class weighting on imbalanced sensor data. Used a stratified train and validation split with computed class weights for the rare failure class, then scored the final model against the separate held-out test set that ships with the dataset. The implementation is refactored into modules for config, preprocessing, models and evaluation rather than living in a single notebook, with seeded runs for reproducibility.

Key Result

Model 4 (Adam optimizer, 14→7 architecture) is recommended for deployment — on held-out test data it catches 85% of true generator failures at 97% precision on flagged failures, the strongest recall/precision balance among the 7 configurations tested.