A Multivariate Time Series Anomaly Detection Model Based on Wavelet Decomposition for Spatiotemporal Feature Fusion

Regarding the task of anomaly detection in multivariate time series, the signals collected by different sensors are diverse. For instance, there are vibration signals with drastic frequency changes and temperature signals with gentle changes. The data distributions of different sensors are different, and the characteristics they exhibit are also distinct. In response to this, we propose a spatiotemporal feature fusion anomaly detection network based on wavelet decomposition. Firstly, the time series is decomposed into different frequency components, and then for the same frequency, a graph attention network of inter-sequence space and intra-sequence time is employed for feature fusion. In the end, a hybrid prediction network integrating time point prediction and time period segmentation prediction is designed to distinguish anomaly points. Through extensive experimentation, we validated the model's performance on various general datasets.