Application of Transformer Based on Escape Optimization Algorithm to the Faulty Bearing Classification Problem

Deep learning has shown strong feature extraction and classification abilities in fault diagnosis, especially for vibration signal time series. The Western Reserve bearing dataset is a common benchmark for evaluating diagnostic models. This paper proposes a Transformer model optimized with the Escape Algorithm (ESC) to enhance bearing fault feature extraction and classification accuracy. Innovations include a multi-scale convolution-attention fusion module for better local temporal feature extraction and the first use of ESC for Transformer hyperparameter optimization. Experiments on the Western Reserve dataset show the model achieves 96% accuracy on both training and test sets, with improved robustness in distinguishing ambiguous categories. The results demonstrate the model’s effectiveness for complex fault recognition, offering valuable insights for intelligent manufacturing and equipment health monitoring.