Predictive maintenance for conveyor drive bearings from vibration data
Use bearing vibration data to build features, a fault classifier and an anomaly detector, design alarm logic with a known false-alarm rate, and show the downtime it saves.
Industry project · for Al Khaleej Cargo Terminal, an air-cargo handling terminal in Dubai South running 140 belt-conveyor drive motors in a 24-hour operation
Al Khaleej moves about 1,900 tonnes of air cargo a day. Its baggage-style conveyors are driven by 140 induction motors of 2.2 to 7.5 kW, most running at about 1,450 rpm. Last year 11 drive bearings failed without warning. Each failure stopped a conveyor line for an average of 5.5 hours, delayed flights' cargo loading, and cost an estimated AED 38,000 in penalties, overtime and parts.
Maintenance today is a monthly walk-round with a handheld vibration meter. It catches some problems, but bearing damage can go from first sign to failure in two or three weeks, and summer heat in the conveyor halls, often above 40 °C, speeds wear. The reliability manager wants wireless vibration sensors on every drive, with software that spots developing bearing faults early and does not cry wolf.
Nobody at the terminal has built this before, and the sensors are not yet installed. You will use public, labelled bearing vibration data from the Case Western Reserve University (CWRU) bearing data centre, which has healthy and faulty bearings at different loads and fault sizes, to design and test the method.
You will: - study the failure modes and write a sensor plan, including bearing defect frequencies; - process signals with the FFT and envelope analysis and build features; - assemble a dataset with a split that does not leak; - train a fault classifier and a healthy-only anomaly detector, and evaluate them honestly; - design alarm logic with persistence rules and a calculated false-alarm rate per month across 140 motors; - write a business case and maintenance workflow.
You need no sensors. All work runs in Python on the public dataset.
Done looks like a feature pipeline, models evaluated on held-out fault sizes or loads, alarm logic with an expected false-alarm rate under 2 per month across the fleet, and a report the reliability manager can use to justify sensors on all 140 drives.
Sensor plan with defect frequencies + signal processing and feature code + dataset builder with split rationale + trained models with evaluation + alarm logic with false-alarm calculation + report with business case and workflow
Stack: Python, NumPy and SciPy (signal), scikit-learn, pandas, Matplotlib, CWRU bearing dataset
- Physics-based features20
Defect frequencies are calculated for the test bearing and envelope spectra show peaks at those frequencies for faulty cases.
- Honest evaluation30
Models are tested on held-out fault sizes or loads with no segment leakage, and results include per-class recall and false-positive rates.
- Useful anomaly detection15
A detector trained on healthy data only flags faults with a stated threshold, and its healthy false-positive rate is measured.
- Alarm design20
Persistence and escalation rules turn model output into alarms, with a calculated false-alarm rate per month across 140 motors under 2.
- Business case and workflow15
The report estimates failures avoided and savings in AED and describes what maintenance staff do on each alarm level.
Pass 60% · Merit 75% · Distinction 90%
Engineering Students
Diploma, B.Tech, B.E., BCA, MCA, or M.Tech students seeking practical experience.
Aspiring Developers
Developers who want to build proof-of-work capstone projects for tech resumes.
Job Seekers
Graduates looking for AI-evaluated code review & verified industry credentials.