predictivemaintenance.app

Apps · Diagnosis

Fault Classifier

Six machine conditions, hundreds of simulated vibration recordings, three machine-learning models and one set of expert rules. Train them on the same features, read the confusion matrices, then test them on a single signal.

Runs in your browserSynthetic vibration data
Which features the Random Forest relies onPermutation importance on the test set

Train the models to see feature importance.

Test one signalNew machine, never seen in training

What is going on

Each simulated machine runs at its own speed with its own sensor gain, resonance and imperfections. Thirteen features are computed from every window: time statistics (RMS, peak, crest factor, kurtosis, skewness), order amplitudes and ratios (1×, 2×, 3×, 4–10×, 0.5×), high-frequency energy, and envelope amplitudes at the bearing outer-race (BPFO) and inner-race (BPFI) frequencies.

  • Expert rules use fixed thresholds an analyst might write: envelope peaks for bearings, harmonics for looseness, 2× for misalignment, a 1× level for unbalance.
  • kNN votes among the most similar training windows. Random Forest averages many decision trees. The neural network has one hidden layer trained with Adam.

Things to try: switch to shuffled windows and watch accuracy rise for the wrong reason. Switch to faster, noisier test machines and watch it fall. Set severity to early faults only and see which classes get confused with healthy.

Read more in Fault classification, Data and feature engineering and Real-world data challenges.