ReneWind Failure Prediction
Predicting wind-turbine generator failure ahead of time from sensor data, so maintenance crews can act before a costly unplanned breakdown instead of after.
The problem
Generator failures are rare compared to normal operation, which is exactly what makes them hard to predict and expensive to miss. A missed failure means an unplanned outage and repair cost; a false alarm means an unnecessary maintenance visit. The two mistakes are not equally costly, so accuracy alone is a misleading target.
Approach
Seven neural network configurations were trained and compared on the minority failure class specifically, since the dataset is heavily imbalanced toward normal operation. Rather than using the default 0.5 probability threshold, the decision threshold was tuned on the validation set to optimize the F2 score — which weights recall higher than precision — reflecting that missing a real failure is worse than one extra inspection.
Validation set (best model, F2-optimal threshold)
| Class | Precision | Recall | F1 |
|---|---|---|---|
| Normal operation | 99.4% | 99.6% | 99.5% |
| Failure — the class that matters | 93.0% | 89.6% | 91.3% |
Carried to the held-out test set: 98.9% overall accuracy, with the failure class scoring 87.9% recall, 92.9% precision, 90.4% F1 — confirming the threshold choice generalized rather than overfitting to validation.
Overall accuracy (98.9%) looks impressive but is inflated by the class imbalance — a model that always predicted 'normal' would already score in the mid-90s. The failure-class recall (87.9%) is the number that actually matters operationally, and it held up on unseen test data.