Multivariable time series anomaly detection based on cross spatial-temporal attention mechanism

In complex industrial production, anomaly detection utilizing real-time multivariable time series data is crucial for ensuring the stability and safety of production processes. Aiming at the numerous issues with current multivariate time series anomaly detection methods, a novel unsupervised approach is proposed, which called multivariable time series anomaly detection based on cross spatial-temporal attention mechanism method, CSTA-MTAD. Firstly, to address the issue of incomplete extraction of potential spatiotemporal features between sequences and variables, the multi-head self-attention mechanism is used to assign distinct attention scores to different time series and variables, thereby capturing the complex spatiotemporal dependencies and realizing the spatiotemporal unity of the real world, culminating in a cross spatial-temporal attention mechanism. Subsequently, the extracted spatiotemporal information is explicitly fused through gated recurrent units to capture richer and more complete potential spatiotemporal information. To address the issue of inadequate performance in anomaly detection methods that rely solely on prediction or reconstruction models, a framework that jointly optimizes both models is utilized for anomaly detection, which can balance the advantages of the two methods and improve the performance of anomaly detection. Lastly, to address the scarcity of data labels in the real world, an unsupervised framework is employed for model training, thereby reducing the reliance on labeled data. Experimental results demonstrate that the proposed method achieves the best performance across three public datasets. The F1 scores on the SMAP, MSL, and SMD datasets are 0.958, 0.949 and 0.958, respectively, with an average F1 score of 0.955. This represents improvements of 5.94%, 2.63%, and 4.27% over the other six best baseline methods, respectively, highlighting its exceptional capabilities in multivariate time series anomaly detection.