Bearing Anomaly Detection Based on Generative Adversarial Network

It is challenging to detect incipient fault of a bearing in a rotating machine being operated under complex conditions. Current intelligent fault diagnosis models require massive historical data of bearing monitoring signals under different health states and the corresponding state labels. However, in some cases, only the samples when the bearing is under healthy condition are available. To deal with the scenario when anomalous samples are absent, this paper proposes a bearing anomaly detection method based on generative adversarial network (GAN). Specifically, only the normal samples are used to train the proposed GAN model, by which the signal data distribution of normal samples is learned. Then, the newly acquired signals with unknown health states are taken as the input data of the trained model. Residual loss of each new sample is calculated by comparing the dissimilarity between the generated sample and the input sample. Because the trained model is optimized by the normal samples, the residual losses of the anomalous samples will be large due to the different data distribution from that of the normal samples. Finally, the bearing anomaly is detected when the residual loss of newly acquired sample increase quickly during continuous inspection of the bearing health condition. Life-circle vibration signals of a bearing measured from a run-to-failure test are used to validate the proposed method. The result indicates that the bearing anomaly can be detected earlier by the proposed GAN model than the traditional method based on signal time-domain statistical criterion, which is of great significance for incipient weak fault detection of mechanical systems.