Adaptive Anomaly Detection in Industrial Systems: An EVT-DTS Approach with LSTM Autoencoders

The reliability of modern industrial systems is rigid; therefore, it is mandatory to monitor the system status and detect anomalies accurately. A long short-term memory (LSTM) network can be used to predict the trend of single-dimensional system test data and implement anomaly detection based on the prediction result. However, the structure of the modern system is complex, and strong dependencies may exist between different variables. The LSTM-based detection method cannot capture this dependency, and some anomalies can be ignored. Therefore, an anomaly detection framework based on the LSTM autoencoder is proposed in this paper. The auto encoder is applied to find the hidden dependency among variables by minimizing the reconstruction error of normal data, while the LSTM is used to capture the temporal dependencies in the time series. Moreover, a new dynamic error threshold selection strategy based on extreme value theory (EVT-DTS) is presented, which can avoid estimating the error distribution beforehand. The EVT-DTS method can dynamically adjust the error threshold according to the current input data error so that the overall optimal detection result can be obtained. Finally, we implement experiment on two industrial applications using the proposed method, which demonstrate its effectiveness in finding complex anomaly states in the system.