Multi-Scale Deep Convolutional Time Series Anomaly Detection Based on Correlation Matrix

With the proliferation of multivariate time series data in various fields, this paper proposes an anomaly detection method based on modelling the correlation of multivariate time series. Multiscale Deep Convolutional Anomaly Detection Model (MSDCAD) based on a correlation matrix is introduced. It characterizes system states at different time steps by constructing multiscale correlation matrices and encodes inter-series correlation information using a deep convolutional encoder. Secondly, temporal patterns are captured using a convolutional long short-term memory network with an attention mechanism, and the correlation matrix is reconstructed through a convolutional decoder. Finally, anomaly detection is conducted by analyzing the residuals between the original and reconstructed correlation matrices. Testing on multiple datasets has proven the effectiveness of the MSDCAD model in detecting anomalies in multidimensional time series.