An Improved Variational Autoencoder-Based Anomaly Detection Model for Rolling Bearing

In practical industrial scenarios, engineering equipment typically operates under normal monitoring conditions, where acquired signals lack fault data. Moreover, it is difficult to extract the fault features of the nonlinear vibration signal obtained by the sensor and the calculation cost is high. In response to this challenge, this work introduces an anomaly detection method based on a Variational Autoencoder (VAE) improved with the assistance of a Generative Adversarial Network (GAN). This proposed method introduces adversarial generation ideas into the VAE framework and uses only normal bearing samples during training to establish an anomaly score threshold. By setting an anomaly status threshold, the method effectively determines whether a bearing has developed a fault. The accuracy of this method is further verified through practical applications and validation on the publicly available XJTU-SY bearing dataset. Comparative analyses are conducted using three metrics: early fault detection location, false alarms, and missed detections to evaluate the performance of the anomaly status detection. The experiments demonstrate that the VAE-GAN outperforms traditional data-driven algorithms in terms of early fault detection performance.