A Frequency-Enhanced Transformer Model and Its Application in the Fault Diagnosis

In recent years, deep learning-based intelligent fault diagnosis methods have made significant progress. However, most of these approaches rely heavily on large amounts of training data and involve substantial computational costs, which can adversely affect both model performance and real-time applicability. To achieve high accuracy with low computational overhead in end-to-end fault diagnosis, this paper directly utilizes one-dimensional vibration signals as model inputs and introduces frequency domain analysis into the Swin Transformer framework, proposing a novel Frequency-Enhanced Transformer method. Specifically, we design a Frequency Enhanced Block that performs frequency domain transformation and data augmentation on one-dimensional vibration signals. In this process, the Fourier-transformed vectors are reorganized into a two-dimensional matrix, providing a richer and more informative input representation. Additionally, a new frequency-enhanced attention mechanism is developed to replace the traditional attention mechanism during model training. Building upon these frequency domain data augmentation and attention enhancements, the network architecture is further optimized. Experimental results demonstrate that, compared to the standard Swin Transformer using time-domain signals, the proposed method reduces computational cost by approximately 9% while achieving 99.225% accuracy. Moreover, our approach exhibits faster convergence and superior generalization capabilities across cross-domain tasks. Finally, when applied to real-world industrial motor data, the method achieves 89.2% accuracy, highlighting its practical value for industrial fault diagnosis.