Visual explanation of neural network based rotation machinery anomaly detection system

Abstract

To make a practical anomaly detection system for rotating machinery in large infrastructures, such as wind turbines, providing an explanation along with the detection results is important so that faults can be easily verified by human experts. Therefore, a method for providing a visual explanation of the predictions of a convolutional neural network (CNN)-based anomaly detection system is considered in this paper. More specifically, the CNN used takes the monitoring target machine’s vibrational data as input and predicts whether the target’s state is healthy or anomalous. A CNN visualization technique is applied this network to obtain an explanation of its predictions. In order to evaluate the obtained explanation, it is compared with an expert diagnosis made on the same data set. The results indicate that the frequency used by the experts to detect faults was also included in the network’s explanation, indicating that the proposed visualization method can be used to provide useful information to help experts verify faults.