Rotating machinery such as induction motors, rolling element bearings, and gearboxes form the backbone of modern manufacturing and process industries. Unplanned machinery downtime due to bearing or gear faults can result in significant production losses, safety hazards, and costly maintenance expenditures. Traditional threshold-based vibration monitoring methods suffer from limited diagnostic resolution in noisy industrial environments and require expert domain knowledge for interpretation. This study presents a systematic comparative investigation of six machine learning classifiers — Support Vector Machine (SVM), Artificial Neural Network (ANN), Random Forest (RF), K-Nearest Neighbour (KNN), one-dimensional Convolutional Neural Network (1D-CNN), and Long Short-Term Memory (LSTM) — applied to vibration signal-based fault diagnosis in rolling element bearings. The Case Western Reserve University (CWRU) Bearing Data Centre dataset, comprising vibration signals from normal, inner race fault, outer race fault, and ball fault bearing conditions at four load levels, was used for experimental validation. Time-domain, frequency-domain, and time-frequency domain (Wavelet Packet Transform) features were extracted and evaluated. Results demonstrate that the 1D-CNN achieves the highest classification accuracy of 97.1% at SNR = 10 dB and 94.2% at SNR = 5 dB, outperforming traditional handcrafted-feature-based classifiers. Feature importance analysis using SHAP values reveals that kurtosis, root mean square, and spectral centroid are the most discriminative features for bearing fault classification. The proposed framework provides a deployable solution for real-time condition monitoring in industrial environments with limited computational resources.
Rajiv Kumar, Sanjay Bose, Neha Chaudhary (Thu,) studied this question.