# TSA-Net: Multivariate Time Series Anomaly Detection Based on Two-Stage Temporal Attention

## TL;DR

TSA-Net is a new method for detecting anomalies in multivariate time series that improves accuracy and reduces training time for industrial monitoring.

## Contribution

TSA-Net introduces a two-stage spatio-temporal attention framework with a reparameterized architecture for efficient training and deployment.

## Key findings

- TSA-Net improves the F1 score by approximately 7% compared to existing methods.
- The model reduces training time by up to 99% compared to complex Transformer-based models.
- Experiments on three benchmark datasets show TSA-Net's effectiveness in high-dimensional anomaly detection.

## Abstract

Multivariate time series anomaly detection is a critical technique for industrial intelligent monitoring. However, existing methods often suffer from prohibitively high training costs and slow convergence, making them ill-suited for industrial scenarios that require frequent model retraining due to dynamic operating conditions. To this end, an efficient two-stage spatio-temporal attention detection framework, TSA-Net, is proposed. This framework adopts a two-branch architecture utilizing a structurally reparameterized temporal convolutional network (RepVGG-TCN) and a graph attention network (GAT). Crucially, the RepVGG design enhances feature extraction capability during training through a multi-branch structure while collapsing into a compact single-branch architecture for deployment, thereby optimizing structural complexity. At the core of TSA-Net is a cascading feedback mechanism, where preliminary predictions from the first stage serve as guidance signals to augment the input for the second stage, enabling coarse-to-fine iterative refinement. Furthermore, an adaptive gating mechanism dynamically fuses spatio-temporal features, improving the model’s adaptability. Extensive experiments with ten state-of-the-art algorithms on three benchmark datasets demonstrate that TSA-Net achieves significant optimization. Specifically, it improves the F1 score by approximately 7% while reducing the training time by up to 99% compared to complex Transformer-based models, offering a rapid-deployment solution for high-dimensional anomaly detection.