CLS GAN: Integrating Autoencoders and Transformers for Enhanced Bearing Fault Diagnosis

Abstract

Bearing fault diagnosis with limited samples is a key challenge in the field of intelligent manufacturing, necessitating the development of models capable of accurate learning from constrained data with strong generalization capabilities. This study proposes a novel framework combining autoencoders and generative adversarial networks, termed the Conditional Latent Space Generative Adversarial Network (CLS GAN), which utilizes autoencoders to learn the latent data distribution of signals, effectively capturing and reproducing the complexity of fault signals. Enhanced with an improved Transformer structure, this model is able to process and recognize long temporal features between signal segments, thereby boosting the accuracy and efficiency of fault diagnosis. Through the architecture of a Conditional GAN, a multi-class task discriminator is implemented, enabling effective fault type discrimination under conditions of limited samples. In situations where samples are restricted, the proposed CLS GAN model achieved an accuracy of 75% on the CWRU dataset, demonstrating the efficacy and practicality of an integrated framework that combines advanced generative adversarial networks and Transformer technology in mechanical fault diagnosis.