<scp>DPAC</scp> ‐ <scp>FAT</scp> : A Lightweight Dual‐Path Adaptive Convolution and Frequency‐Aware Transformer for Multivariate Time Series Anomaly Detection

ABSTRACT Multivariate time series anomaly detection (MTSAD) is widely applied in domains such as industrial monitoring and financial risk management. Although unsupervised deep learning models have made notable progress in capturing normal behaviour patterns, they still face a trade‐off between modelling capacity and computational efficiency. Transformer‐based methods possess strong representational power but are computationally intensive, while lightweight models are more efficient but often struggle to capture fine‐grained anomalies such as periodic perturbations and frequency shifts. To address this issue, we propose a unified, lightweight, and highly discriminative framework, Dual‐Path Adaptive Convolution and Frequency‐Aware Transformer (DPAC‐FAT), which jointly models local disturbances and frequency‐domain anomalies. Specifically, a Structure‐Aware Temporal Encoder (SATE) is introduced to efficiently model nonlinear short‐term dynamics, while a Frequency‐Informed Contextual AutoEncoder (FICA) is employed to enhance the recognition of periodic structures and frequency‐specific variations. Considering that the temporal and frequency branches should exhibit consistent representations under normal conditions, we further design a Dual‐Branch Alignment Loss (DBAL) based on cosine similarity to encourage both pathways to learn coherent backbone representations, thereby improving model robustness and cross‐view discriminability. Experimental results on five real‐world datasets demonstrate that DPAC‐FAT achieves a 5.41% improvement in average F1 score over 11 representative state‐of‐the‐art methods, while significantly reducing training costs.