Reconstruction-based Multi-Scale Anomaly Detection for Cyber-Physical Systems

This paper considers anomaly detection for cyber-physical systems, in which the multivariate time series data collected from different sensors have complex temporal dependencies and inter-sensor correlations. We firstly propose an improved unsupervised anomaly detection framework which extracts the temporal and spatial patterns based on the autoencoder and the attention-based convolutional long-short term memory networks. In particular, the original data are fused into the input signature matrices to avoid information loss and an improved sample-based threshold setting approach is proposed to estimate the optimal threshold automatically. Finally, the experiments on two sensor datasets illustrate that our model achieves superior performance over state-of-the-art methods.