Fault Diagnosis of Ball Bearing Induction Motors for Predictive Maintenance (www.mdpi.com)

This paper presents a practical machine‑learning pipeline for fault diagnosis of ball bearings in induction motors aimed at predictive maintenance. The method extracts statistical and time–frequency vibration features (V_RMS, a_RMS, a‑Peak, Crest Factor, temperature, STFT/Wavelet/CWT/PSD-derived descriptors), reduces dimensionality (PCA/KPCA), then classifies faults with compact neural networks (input sized to PCA output, 1–2 ReLU hidden layers, Softmax/Sigmoid output) and SVMs. Experimental validation on both lab data and public datasets (CWRU, MFPT) shows high diagnostic performance—around 99% accuracy for the proposed ANN/SVM pipeline and 98.7% for comparable CWT+ANN/SVM approaches—while demonstrating robustness to noise and load variation. Results note ANN edge in complex, nonlinear fault patterns and SVM’s advantage for low‑compute/edge deployment. For the AI/ML community this work underscores that carefully engineered feature extraction + compact models can match more complex deep architectures for industrial PdM, offering lower latency and easier edge integration. The paper also outlines next steps relevant to research and deployment: multimodal sensor fusion (acoustic, current, temperature), automated feature selection (genetic algorithms, autoencoders), leveraging unlabeled data, and exploration of ViTs, GNNs, attention and LSTM/CNN hybrids for richer temporal/spatial representations. Practical integration (LabVIEW, k‑fold validation, SHAP/permutation importance) is emphasized for trustworthy, real‑time industrial use.