AnomalyTCN: Efficient contrastive-based time series anomaly detection with pure convolution structure

In recent years, contrastive-based time series anomaly detection has emerged rapidly and become compelling thanks to its performance superiority. But at the same time, we observe that existing contrastive-based methods can only work with the complicated attention mechanisms, which brings heavier computational costs and less framework generality, preventing their further promotion. To handle these challenges, we pioneer an attention-free contrastive-based solution in this paper. Through in-depth analysis, we find that contrastive discrepancy learning doesn't only work with attention. By designing asymmetric dual-branch convolution structure and adopting asymmetric supervision training strategy, a simple and light-weight pure convolution structure can also produce effective contrastive representations for anomaly detection. Based on above findings, we propose AnomalyTCN as a more efficient and effective solution for contrastive-based time series anomaly detection. By adopting a simple and light-weight pure convolution structure to avoid the complicated attention computation, our method can greatly improve the efficiency while maintaining the performance superiority of contrastive-based methods. Experimentally, our AnomalyTCN achieves the consistent state-of-the-art performance on various time series anomaly detection tasks while saving 83.6% running time and 20.1% memory usage. These results validate that our AnomalyTCN is an ideal solution for time series anomaly detection with a better balance of performance and efficiency. Meanwhile, our study also reveals the possibility of combining contrastive-based anomaly detection frameworks with other efficient time series backbones beyond the costly attention mechanism, demonstrating the great potential in this novel research direction.