Unsupervised Anomaly Detection Method for Bearing Based on VAE-GAN and Time-Series Data Correlation Enhancement (June 2023)

Abnormal detection (AD) plays a crucial role in maintaining system integrity and minimizing risks associated with abnormal behavior or events. Given the practical challenges in obtaining data that covers the entire lifespan of mechanical equipment, we proposed a novel unsupervised AD method with a high sensitivity to early failures. In the proposed time convolutional variational autoencoder - generative adversarial networks (TCVAE-GAN) model, the temporal convolutional network (TCN) module is incorporated into the encoder, decoder and discriminator to improve the reconstruction and feature extraction capabilities of time-series data. Subsequently, the implementation of adversarial training leads to the continuous enhancement of the reconstruction ability of the variational autoencoder and the improvement of the discriminator’s capability to effectively detect early weak faults. Furthermore, utilizing the output of the discriminator directly as a healthy indicator (HI) for AD enhances the interpretability of the results. To verify the accuracy and effectiveness of the proposed method, we conduct a comparison with traditional time-domain methods and five other unsupervised methods. The validation results obtained from both the laboratory datasets and actual wind turbine datasets demonstrate the high sensitivity of our method in detecting early weak faults in bearings.