Anomaly detection on rotating machines with AI

Context

Three labs of the AI module in the first-year Master's MESI, supervised by Ivan Vazquez-Munoz and Dominique Knittel, in March 2026. The same set of six signals is used in every lab, with two healthy states and two families of faults. Each lab switches to a different learning family.

What I did

• In the unsupervised lab, I extracted nine time-domain features, then tested K-means, DTW, One-Class SVM and Isolation Forest.
• In the supervised lab, I tested KNN, CNN and LSTM, with overlapping windows to augment the data.
• In the semi-supervised lab, I trained an autoencoder on the healthy signal only, on spectrograms then scalograms.

What I noticed and learned

• Simple and robust methods suit plant monitoring. The scalogram helps trace a fault back to its origin.
• No method is enough on its own. Cross-checking the results gives confidence.

My mistakes

• The dataset is too small for large models (125 windows). The CNN and the LSTM overfit quickly.
• The Isolation Forest missed the mild faults. Tuning its contamination is still to be done.

Results

• The One-Class SVM is the most rigorous of the unsupervised methods.
• Window overlap up to 75% gives a clear gain.
• The autoencoder separates healthy and faulty signals well, except for very short shocks.
• A simple KNN is often more reliable than the deep models.

Tech stack