36.1-40.0 dB
BPFO prominence
Median outer-race evidence across 3 and 0 HP
Engineering case study
Motor Bearing Condition Monitoring from Vibration Signals
Physics-first vibration analysis using bearing kinematics, calibrated spectra, and Hilbert-envelope demodulation on CWRU data.
- Public-dataset analysis
- Automated test
Problem statement
Distinguish normal, inner-race, outer-race, and rolling-element conditions across two loads while keeping the evidence physically interpretable and avoiding same-record machine-learning leakage.
A reproducible diagnostic workflow that uses mechanical fault frequencies and demodulation before introducing a deliberately limited cross-load classifier.
34.8-35.0 dB
BPFI prominence
Median inner-race evidence across 0 and 3 HP
5.8-9.9 dB
BSF prominence
Weaker, load-sensitive rolling-element evidence
82.5%
Cross-load accuracy
Exploratory; windows are not independent machines
Project scope
- Eight official CWRU drive-end records covering four conditions at 0 and 3 HP
- Smallest 0.007-inch seeded defects and five one-second windows per record
- 12 kHz common analysis rate, 1 Hz FFT resolution, and a fixed 2-5 kHz envelope band
- Exploratory cross-load logistic regression after physics-based interpretation
System or processing architecture
Checksum-verified MAT files pass through channel discovery, offset removal, rate standardization, and segmentation. A parallel feature path computes time statistics while a spectral path calculates kinematic frequencies, raw spectra, resonance-band envelopes, and local frequency prominence.
My engineering contributions
- Selected a controlled subset and encoded official record metadata, source rates, loads, speeds, and hashes.
- Implemented amplitude-correct one-sided FFTs, bearing kinematics, envelope analysis, and interpretable feature extraction.
- Designed a cross-load experiment that keeps windows from each source record on one side of a directional split.
- Produced the notebook, seven figures, results tables, automated tests, and deployment considerations.
Important design decisions
- Used measured RPM in the kinematic equations rather than a nominal motor speed.
- Fixed the 2-5 kHz demodulation band for all conditions to avoid post-hoc favorable tuning.
- Treated the classifier as secondary because 40 windows still represent only eight source recordings.
Engineering methods
- Polyphase anti-alias resampling and one-second segmentation
- Hann-windowed coherent-gain-corrected FFT
- BPFO, BPFI, BSF, and FTF kinematic calculation
- Butterworth bandpass and Hilbert-envelope spectrum
- Cross-load standardized logistic-regression comparison
Selected evidence
Readable results, with context
Verification and testing
- Official downloads are checked against fixed SHA-256 values before analysis.
- Tests cover rate and length, offset removal, segmentation, FFT amplitude/peak location, Nyquist endpoint, and feature values.
- The analysis produces 40 balanced feature rows and seven publication-quality figures.
- The notebook and two-page report preserve the same assumptions, units, and limitations as the executable pipeline.
Honest limitations
- CWRU faults are isolated EDM-seeded defects on a laboratory rig, not naturally evolving field failures.
- Each class/load combination has one source recording; window count does not equal independent asset count.
- Normal and fault records have different source sampling histories before standardization.
- Absolute acceleration calibration is not established by the referenced public pages, so amplitude remains in CWRU units.
Tools and technologies
- Python
- SciPy
- NumPy
- pandas
- scikit-learn
- CWRU Bearing Data Center
- Hilbert transform
Skills demonstrated
- Vibration analysis
- Spectral estimation
- Envelope demodulation
- Bearing kinematics
- Condition monitoring
- Experimental design
Downloadable proof of work
Review the evidence directly
Continue in Embedded Electronics & Signal Processing