A Difference-Sensitive Mechanism Transformer Model for Incipient Fault Diagnosis in Electrical Drive Systems

Abstract

The wide application of electrical drive systems in industrial and transportation domains necessitates fault diagnosis technology. In the past decade, transformer models have emerged as critical analytical tools, owing to their exceptional temporal sequence processing capabilities and proficiency in detecting intricate operational patterns. However, traditional transformer often struggles to capture minor fault features in the system. To solve this challenge, an incipient fault diagnosis method using transformer with a difference-sensitive mechanism embedded into the broad echo state network framework (TDMB-ESN) is proposed in this article. The advantages of the proposed difference-sensitive mechanism mainly include three parts: 1) an adaptive attention window is proposed to calculate the attention score matrix, which enhances the flexibility of the transformer in extracting local features of fault data; 2) a difference measure function is designed to further expand the distance scale and strengthen the difference between different fault data points; and 3) the softmax function based on the difference measure function (DM-softmax) is designed. The experimental results show that the proposed method can significantly improve the accuracy of incipient fault diagnosis in electrical drive systems.