Time Series Industrial Data Anomaly Detection Algorithm Based on Transformer

Multivariate time series anomaly detection has important application value in fields such as finance, industrial manufacturing, water treatment, transportation logistics, and network security. However, existing methods cannot learn the spatiotemporal correlation between data well. To address these issues, this paper proposes a method based on the combination of improved Transformer and CST-GL, aiming to efficiently mine the spatiotemporal correlation between data and improve anomaly detection performance. Firstly, the Transformer is combined with gradient optimized position encoding to more effectively extract temporal information between data. Secondly, in order to more effectively extract spatial correlation between data, a dual dimensional attention mechanism block is designed in the Transformer to simultaneously extract temporal and spatial correlation information. The self-attention mechanism is integrated into the GNN to dynamically extract spatial correlation between data. Finally, L pre training modules are set up to each head of the Transformer to extract particularly important temporal and spatial information between data. The experimental results demonstrate that the proposed method achieves excellent performance in anomaly detection.