Dual-Branch Transformer with Contrastive Learning for Unsupervised Time Series Anomaly Detection: Special Session: SMII- Intelligent Manufacturing and Industrial Intelligence

Time series anomaly detection is critical for various applications including industrial monitoring, cybersecurity, and healthcare. However, existing methods face challenges in learning robust representations from normal data alone. In this paper, we propose a novel dual-branch transformer architecture with contrastive learning for unsupervised time series anomaly detection. Our method consists of two complementary branches: a reconstruction branch that captures temporal patterns through anomaly-aware attention, and a contrastive branch that learns discriminative representations through temporal augmentation. Extensive experiments on two benchmark datasets (CATS and SMD) demonstrate that our method achieves state-of-the-art performance, with F1 scores of 0.670 and 0.677 respectively, outperforming strong baselines by 9.5% and 3.8%. Ablation studies confirm that each component contributes positively, with temporal augmentation providing the largest gain on largescale data. Hyperparameter sensitivity analysis demonstrates robustness across a wide range of settings, and the dual-branch architecture introduces negligible inference overhead despite increased model capacity.