1
Engineer vibration features by hand
Build the time-domain numbers (RMS, kurtosis, crest factor) and the envelope-spectrum energy at the defect frequencies that cleanly separate a healthy bearing from a faulty one.
Become a German condition-monitoring engineer. A motor that sounds fine will seize in three weeks — hear the crack inside the bearing from a sensor on the outside. You build a real fault detector by hand in Python: engineer the vibration features, train a tiny classifier, locate the fault from its frequency signature, test it on bearings it never saw, and set an alarm threshold on cost. Physics-based synthetic data; not affiliated with or endorsed by Schaeffler.
Meet the failing motor and your coach, and commit to building the detector yourself. No code yet — just the mission and the bar.
Load the faulty bearing snapshot in Python and plot it beside a healthy one. Spot the periodic knock the healthy signal lacks.
Author the time-domain features (RMS, kurtosis, crest factor) by hand and identify which one jumps most on the faulty bearing.
Use the provided envelope-spectrum helper to compute band energy at the defect frequencies, then name the dominant frequency and locate the cracked ring.
Train a shallow, depth-capped decision tree on the training bearings only and state its depth and node count against the embedded budget.
Evaluate the detector on held-out bearings for a confusion matrix, catch-rate, and false-alarm rate, then pick an alarm threshold on cost.
See whether your detector caught the fault with weeks of lead time, then file the assembled pipeline and decision memo for evaluation.
Build the time-domain numbers (RMS, kurtosis, crest factor) and the envelope-spectrum energy at the defect frequencies that cleanly separate a healthy bearing from a faulty one.
Fit a tiny depth-capped decision tree sized to an embedded compute budget, trained leakage-free on the training bearings only, and locate the fault from its frequency signature.
Run an honest held-out evaluation (catch-rate vs false-alarm rate) and defend a cost-based alarm threshold in a one-page maintenance memo.