MaelNet: Memorable Anomaly Election Learning Detecting Anomalies in Time Series Data using Dual-Net Transformer

Detecting anomalies in time series data is critical for identifying unusual patterns across a wide range of applications. This study introduces MaelNet, an innovative Dual-Net Transformer model enhanced with reinforcement learning, designed to address challenges such as data imbalance and the evolving nature of temporal anomalies. MaelNet comprises two fast learners—an Anomaly Transformer and a DC Detector—alongside a slow learner that employs a Modified Non-Stationary Transformer. Importantly, the input to the slow learner differs from that of the fast learners: the time step input is initially masked, allowing the model to capture the inherent bias in time series data. This design enhances anomaly detection by dynamically modeling both biased and unbiased temporal sequences. Evaluated on five real-world datasets (SMAP, SMD, SWaT, PSM, and MSL), MaelNet surpasses existing Transformer-based baselines, achieving an F1-score improvement of 3.84%and a reduction in false positive rate by 1.109%. These findings underscore the effectiveness of MaelNet’s dual-attention architec-ture and adaptive reinforcement learning in capturing complex temporal dynamics.