Feature consistency anomaly detection method based on adversarial training deep autoencoder

Real-time and accurate fault diagnosis is essential to ensure the safe and stable operation of industrial systems. However, most of the monitoring data collected by sensors in actual industrial sites are obtained when the system is operating in a healthy state, so it is difficult to obtain abnormal data with fault labels. Therefore, it is of great practical significance to carry out research on unsupervised anomaly detection during equipment operation. This paper proposed an anomaly detection method based on the consistency of features of adversarial training deep autoencoder. This method fed two random training sets into two deep autoencoder network, and designed loss functions by the error of input and output consistency and the error defined by the degree of feature inconsistency. Then the network parameters are updated by back-propagation through adversarial training of the two deep autoencoder network, and then the anomaly score is obtained by using the weighted sum of the generated and discriminated losses, and finally unsupervised anomaly detection is performed by the discrepancy of the anomaly score. The efficiency of the proposed method is verified by the data set of gearboxes.