Time-Frequency RWGAN for Machine Anomaly Detection Under Varying Working Conditions

Obtaining current fault data for mechanical equipment is a challenging endeavor. Despite some successes in anomaly detection, achieving satisfactory results remains difficult, particularly when dealing with datasets containing few instances of anomalies and significant distribution differences. To address this challenge, a novel deep residual Wasserstein generative adversarial network, named RWGAN, is designed to effectively detect anomalies of unseen samples in rotary machines under varying working conditions. Initially, an encoder-decoder–encoder pipeline is constructed based on the convolutional autoencoder (AE) module to extract deep feature representations from the time-frequency transformation of vibration signals. In addition, the ResNet structure with skip connections is embedded into the model to enhance feature learning and model performance. Furthermore, a Wasserstein distance module is developed, integrating loss-specific feature learning networks and adversarial training techniques to address large distribution discrepancies across data from varying working conditions. Finally, the network is updated in an end-to-end manner to generate real-like output by fitting the probability distribution of time-frequency images. To validate the effectiveness and superiority of the proposed method, three cases across 15 tasks under varying working conditions are designed. The results demonstrate that the proposed approach achieves satisfactory anomaly detection performance and outperforms other state-of-the-art methods.