# Fault detection of high-speed train wheelset bearings based on improved auxiliary classifier generative adversarial networks and VAE

## TL;DR

This paper introduces a new method for detecting faults in high-speed train bearings using improved generative models to handle limited fault data.

## Contribution

A novel supervised generative model combining IACGAN and VAE is proposed to improve fault detection accuracy with imbalanced data.

## Key findings

- The proposed method generates higher quality samples compared to other GAN variants.
- The detection model achieves 88.04% classification accuracy, a 15.17% improvement over existing methods.
- The model effectively addresses accuracy degradation caused by imbalanced fault samples.

## Abstract

Fault detection in high-speed train wheelset bearings is paramount for ensuring operational safety. However, the scarcity of fault samples limits the accuracy of traditional detection methods. To address this challenge, this paper proposes a supervised generative model that integrates an Improved Auxiliary Classifier Generative Adversarial Network (IACGAN) with a Variational Auto-Encoder (VAE). Firstly, the method employs the VAE as the generator, introducing latent variables with prior information to optimize the encoding and generation process; Secondly, an independent classifier network is integrated into the ACGAN framework to enhance compatibility between classification and discriminative capabilities. Concurrently, a loss function incorporating Wasserstein distance and gradient penalty terms is designed to prevent gradient vanishing during training while satisfying Lipschitz constraints, thereby improving model stability. Experiments conducted on the XJTU bearing dataset validate that samples generated by the proposed method demonstrate superior quality assessment at both the data and feature levels compared to several GAN variants. Furthermore, the constructed VAE-IACGAN-CNN detection model achieves an average classification accuracy of 88.04%, representing a maximum improvement of 15.17% over comparative methods. This significantly mitigates accuracy degradation caused by sample imbalance, demonstrating the proposed approach’s efficacy in resolving low fault detection accuracy stemming from imbalanced high-speed train wheelset bearing samples.