Digital twin of an induction motor and anomaly detection

Context

Second-year Master's project in MESI, ongoing. In Simulink, I am building a digital twin of an induction motor, healthy and then with faults, to see what the stator current reveals about them. The latest summary (6 October 2026) covers a gear fault.

What I did

• Model variants with faults (broken bars, eccentricity, bearing, stator imbalance, gear, torque surges, wiring).
• Parallel simulations, current analysis by FFT and spectrogram.
• Physical, statistical and EMD features over 1 s windows.
• 12 one-class detectors trained on healthy operation only, conformal threshold at 5%.

What I noticed and learned

• Physical features win for every model. Generic statistics detect nothing.
• With 72 training windows, an autoencoder does worse than a Mahalanobis distance.

My mistakes

• I picked the best configuration on the test set, so the scores are slightly optimistic.
• At 1% noise, no model beats a single well-chosen feature (AUC 0.92).
• Everything still comes from simulation, real test bench to come.

Results

• Ongoing project, 230 experiments (12 models, 5 feature sets, 2 noise levels), 0 failures.
• Best model ECOD, with no hyperparameters, mean AUC 0.994 at 0.2% noise and 0.815 at 1%.
• Fault detected 90% of the time from an amplitude of 0.01 p.u. with 0.2% noise.

Tech stack