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.
Explore
Predictive Maintenance: Industrial ML for Fault Detection
ProoV
4.4· 49 ratings
About this project
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.
Ideal for entry-level careers in
Add this to your LinkedIn profile
Projects
inPredictive Maintenance: Industrial ML for Fault Detection
Just completed
Associated with ProoV
A graded ProoV project, completed step by step and marked against a published rubric.
❖condition-monitoring, vibration-analysis and more skills
What you'll do
1
Cold Open & Onboarding
Meet the failing motor and your coach, and commit to building the detector yourself. No code yet — just the mission and the bar.
2
Act 1 · Hear the Failure
Load the faulty bearing snapshot in Python and plot it beside a healthy one. Spot the periodic knock the healthy signal lacks.
3
Act 2 · Turn a Wiggle Into a Number
Author the time-domain features (RMS, kurtosis, crest factor) by hand and identify which one jumps most on the faulty bearing.
4
Act 3 · Find the Fingerprint
Use the provided envelope-spectrum helper to compute band energy at the defect frequencies, then name the dominant frequency and locate the cracked ring.
5
Act 4 · Teach the Sensor to Decide
Train a shallow, depth-capped decision tree on the training bearings only and state its depth and node count against the embedded budget.
6
Act 5 · The Cost of Being Wrong
Evaluate the detector on held-out bearings for a confusion matrix, catch-rate, and false-alarm rate, then pick an alarm threshold on cost.
7
Reveal & Debrief
See whether your detector caught the fault with weeks of lead time, then file the assembled pipeline and decision memo for evaluation.
What you'll learn
2
Train a compact, explainable detector
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.
3
Make the call on cost, not accuracy
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.
What students say
This project gave me practical experience in predictive maintenance and industrial machine learning. I learned how to turn vibration signals into useful features, identify bearing faults using frequency signatures, train a lightweight decision tree, evaluate it on unseen data, a…
I understood how industries process the signals and then extract the essential information for evaluation and also how to evalute the datat
I have gained practical hands on experience and industrial machine learning. I learned how to process raw vibration signals to extract key statistical features (RMS, Kurtosis, Crest Factor), pinpoint bearing faults using frequency spectrum analysis (BPFO), build decision tree mo…
I learned how to turn raw machine vibrations into edge ML features, optimize sensor thresholds for minimal plant costs, and isolate a real-world bearing faults.
This project improved my understanding of predictive maintenance, vibration analysis, feature extraction, and TinyML. I learned how to build, evaluate, and tune a machine learning model using real engineering metrics such as catch rate and false-alarm rate, and how to explain te…
This project gave me hands-on experience in predictive maintenance using machine learning. I learned how to engineer vibration features, train an embedded-friendly Decision Tree model, identify an outer-race bearing fault using frequency-domain analysis, evaluate the model on un…