A comparative study of machine and deep learning models for time-series-based bearing fault diagnosis of induction...

Accurate fault detection in induction motors (IMs) under varying load conditions remains a critical challenge in industrial condition monitoring (CM). Inspired by the foundational work, which highlighted the impact of mechanical load on fault signature detectability. This study proposes a multi-modal signal analysis approach to bearing fault diagnosis using stator current, rotor speed, and flux-induced voltage signals. A custom fifteen-class dataset was collected, comprising healthy and faulty motor states at 0%, 50%, and 100% load levels. Unlike conventional approaches that rely on extensive preprocessing and handcrafted feature extraction, the proposed framework operates directly on raw signals, enabling a lightweight, computationally efficient, and easily deployable solution. This design significantly reduces implementation complexity while maintaining high diagnostic performance, making it suitable for real-time and industrial applications. Two types of models were evaluated in this study: traditional machine learning models and deep learning models. Experimental results demonstrate significant performance gains compared to single-sensor models, highlighting the benefits of cross-domain signal fusion. Models specifically designed to process time-series data, such as the Temporal Convolutional Network (TCN) and particularly the Long Short-Term Memory (LSTM), exhibit outstanding performance. During the training and validation phases, the LSTM model achieved perfect classification accuracy (100%), outperforming all other evaluated models. However, during the deployment-oriented evaluation on unseen test data, the SVM and TCN models demonstrated the most consistent generalization performance, achieving perfect prediction results across all tested samples. Recent architectures, such as the Transformer, also demonstrate strong potential; with careful hyperparameter tuning, their performance especially in terms of generalization can be further enhanced.