Optimizing Bearing Fault Detection: CNN-LSTM with Attentive TabNet for Electric Motor Systems

Abstract

Electric motor-driven systems are core components across industries, yet they’re susceptible to bearing faults. Manual fault diagnosis poses safety risks and economic instability, necessitating an automated approach. This study proposes FTCNNLSTM (Fine-Tuned TabNet Convolutional Neural Network Long Short-Term Memory), an algorithm combining Convolutional Neural Networks, Long Short-Term Memory Networks, and Attentive Interpretable Tabular Learning. The model preprocesses the CWRU (Case Western Reserve University) bearing dataset using segmentation, normalization, feature scaling, and label encoding. Its architecture comprises multiple 1D Convolutional layers, batch normalization, max-pooling, and LSTM blocks with dropout, followed by batch normalization, dense layers, and appropriate activation and loss functions. Fine-tuning techniques prevent overfitting. Evaluations were conducted on 10 fault classes from the CWRU dataset. FTCNNLSTM was benchmarked against four approaches: CNN, LSTM, CNN-LSTM with random forest, and CNN-LSTM with gradient boosting, all using 460 instances. The FTCNNLSTM model, augmented with TabNet, achieved 96% accuracy, outperforming other methods. This establishes it as a reliable and effective approach for automating bearing fault detection in electric motor-driven systems.