1. Introduction
Unplanned induction motor downtime in industrial plants is frequently traceable to bearing defects, rotor bar breakage, or shaft misalignment, all of which produce characteristic vibration signatures that can be detected before catastrophic failure with appropriate feature extraction and classification.
2. Methodology
Vibration data was collected from a motor test rig under four seeded fault conditions, outer-race bearing fault, inner-race bearing fault, broken rotor bar, and shaft misalignment, plus a healthy baseline, using a tri-axial accelerometer. Eighteen time-domain, frequency-domain and wavelet-packet energy features were extracted per sample and used to train a multi-class SVM classifier with an RBF kernel, tuned via grid search over five-fold cross-validation.
3. Results
The SVM classifier achieved 97.8 percent overall classification accuracy across the five classes, with the lowest per-class recall of 95.1 percent for distinguishing inner-race from outer-race bearing faults, which share overlapping spectral signatures. This outperformed a k-nearest-neighbour baseline (92.2 percent) evaluated on the identical feature set.
4. Conclusion
Wavelet-packet energy features combined with SVM classification provide a reliable, computationally light approach for induction motor fault diagnosis suitable for embedded condition-monitoring hardware. Future work will validate the approach on field data from operating plant motors rather than a controlled test rig.
References
[1] Nandi S. et al., Condition monitoring and fault diagnosis of electrical motors, IEEE Transactions on Energy Conversion, 2005. [2] Cortes C. and Vapnik V., Support-vector networks, Machine Learning, 1995.