Industrial Fault Classification based on Vibration Signals: A Comparative Study between Transformers and Classical Machine Learning Models

Industrial fault classification based on vibration signals plays an important role in modern predictive maintenance strategies, ensuring equipment reliability and minimizing costly operational downtimes. This study investigates the application of deep learning models for vibration-based industrial fault classification. Specifically, we conduct a comparative analysis evaluating the performance of the emerging Transformers architecture against established classical machine learning models widely applied in the literature, namely Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), and Recurrent Neural Networks (RNN). The empirical dataset utilized in this research consists of time-series data acquired from two gyroscopic sensors coupled to a rotary test bench. This experimental setup was subjected to four distinct operational configurations to simulate varying mechanical fault conditions. To extract relevant features and enhance model learning, data pre-processing was rigorously conducted in the frequency domain. The predictive efficacy and robustness of each model were systematically evaluated using the F1-Score metric. The empirical results reveal a notable trade-off between computational efficiency and classification accuracy. Although the Transformers model exhibited a significantly faster training convergence rate, the RNN architecture ultimately achieved the highest overall predictive performance for this specific experimental framework. This research contributes to the field by demonstrating the superior adequacy of RNNs for handling lower-dimensionality structured time-series data in fault diagnosis tasks. Furthermore, it establishes critical theoretical and practical foundations for future research regarding the adoption and optimization of Transformers when dealing with highly complex, large-scale multivariate datasets in industrial environments.