Context
Three lab sessions of the AI module in my first master's year (MESI), supervised by Ivan Vazquez-Munoz and Dominique Knittel, in March 2026. The common thread stays the same across the three sessions, spotting anomalies on the signals of a rotating machine, which is the heart of predictive maintenance. I work on the same set of six signals, A to F, with two healthy states (A and E) and two fault families (B and D, then C and F). Each session tackles the problem with a different learning family, and everything is done in MATLAB.
Approach / Solution
TP1, unsupervised learning. I first extract nine time-domain features from the signals (Shannon entropy, RMS, and others), then I run them through several methods to see which ones separate the states well. I compare a K-means, a DTW distance, a One-Class SVM and an Isolation Forest on the same data. TP2, supervised learning. I train and compare a KNN, a CNN and an LSTM on those features. I then add overlap between time windows, a way to artificially grow the amount of data. TP3, semi-supervised learning. I turn the signals into images, FFT spectrograms then wavelet scalograms, and I train an autoencoder on the single healthy signal (A). A signal whose reconstruction error passes a threshold is flagged as anomalous.
Results
One finding comes back across the three labs, the dataset is too small for the heavy models. With 125 windows, and only 49 spectra for the autoencoder, the CNN and the LSTM overfit fast. A plain KNN often stays more reliable, and no single method is enough. Cross-checking several models is what gives confidence in the verdict. In unsupervised, the One-Class SVM was the most rigorous. Isolation Forest missed the mild faults (B and D) while catching the clear ones (C and F). The DTW distance gives a very readable visual picture, but it is expensive to compute. In supervised, window overlap acts as data augmentation, with a clear gain up to 75% overlap. In semi-supervised, normalizing the spectrograms between 0 and 1 stabilized training a lot. The reconstruction error separates the healthy signal from the faults well, except for very brief shocks that get drowned in the average. I draw an engineer's rule from it, the right tool depends on the goal. To automate monitoring on a factory floor, I keep simple and robust methods, a KNN on features or an autoencoder on the FFT spectrogram. To trace the physical origin of a fault (bearing, gear or motor), I switch to the wavelet scalogram, heavier but far more readable to the eye. Deciding to change a part before a production stop is often worth more than chasing the last point of accuracy. What I could still dig into, getting more data since that is the real bottleneck, trying a variational autoencoder, tuning the Isolation Forest contamination, and combining the time, frequency and wavelet domains in a single diagnosis.