Robust unsupervised anomaly detection via multi-time scale DCGANs with forgetting mechanism for industrial multivariate time series

Abstract

Detecting anomalies in time series is a vital technique in a wide variety of industrial application in which sensors monitor expensive machinery. The complexity of this task increases when heterogeneous sensors provide information of different attributes, scales and characteristics from the same machine. Actually, the challenges of anomaly detection for industrial time series are to design effective pre-processing, feature extraction and overcome the lack of abnormal samples. Recent deep learning models have shown prominent abilities on raw multivariate time series, alleviating these previous works. In this work, a novel framework named multi-time scale deep convolutional generative adversarial network (MTS-DCGAN) is proposed to deal with anomaly detection of industrial time series. Firstly, multivariate time series are transformed into the multi-channel signature matrices via sliding window based cross-correlation computation, and therein forgetting mechanism is introduced to effectively avoid false alarms due to excessive influence of old sequences. Then the framework conducts an unsupervised adversarial training of multi-channel signature matrices and capture their hidden features via deep convolution structure. Besides, a novel threshold setting strategy is proposed to optimize anomaly detection performance under imbalance of normal and abnormal data. Finally, the proposed framework is assessed against the experiments on four datasets. Within this study, experimental results show our framework outperforms the comparison algorithms in terms of model performance and robustness, providing an effective anomaly detection method for industrial multivariate time series.