Turbofan Predictive Maintenance

3-layer Transformer encoder · 50 synthetic engines · dual-head anomaly + RUL  |  View source

Test Set Performance

Loading…

Total Cost
—
r·FN + FP
Recall
—
FN: —
Precision
—
FP: —
RUL RMSE
—
cycles

Plots

Precision-recall curve with operating point
Precision-recall curve (validation set). Loading…
Confusion matrix for anomaly detection
Confusion matrix. Loading…
RUL prediction scatter plot
Predicted vs. true Remaining Useful Life across the test set. Diagonal indicates perfect prediction.

Architecture

Encoder — 3-layer Transformer, 4 attention heads, sinusoidal positional encoding
Input — sliding windows of 50 timesteps × 5 sensor features
Pooling — global average pool over sequence dimension
Heads — binary anomaly classification (focal loss γ=2) + RUL regression (MSE × 0.1)
Training — joint loss, 10 epochs, Adam lr=0.0003