Anomaly detection of multivariate industrial sensing data based on graph attention network

Recent development of Industrial Internet produces a large number of sensor data that record the production status of equipment in the field of industrial intelligence. This paper focuses on how to detect abnormal event from multivariate time series data. We propose an anomaly detection model DWGAT, that first uses DeepWalk method to generate an embedding vector for each sensor node that can represent sensor's behavior characteristics according to prior information, and then establishes a graph describing the initial relationships between sensor nodes. The graph attention network is used to mine and aggregate the data features to predict the multivariate data at the next moment. Finally, the predicted value is compared with the actual observed value, and a calculation method of anomaly score is designed to judge whether there are anomalies in the system according to whether the anomaly score is larger than the threshold value learned by the model. Experiments are established on two real-world public industrial datasets and compared with two baseline models. The results show that the F1 scores of the proposed anomaly detection model on the two data sets are higher than those of the current optimal model.