MS-TDNet: Multi-Scale Feature-Based Network for Time Series Anomaly Detection in AIOps

To address the complexity and dynamic challenges of anomaly detection in multi-source heterogeneous time series data in AIOps, this study proposes a deep network model based on multi-scale feature fusion, named MS-TDNet. It aims to solve key problems in time series within cloud service environments, such as distribution drift, multi-channel redundancy interference, and noise coupling. By innovatively introducing a Discrete Cosine Transform (DCT)-based time series decomposition method, the original signal is decoupled into seasonal and trend components. A Seasonal Multi-Period Forecasting module (SMP) and a Trend Multi-Scale Forecasting module (TMS) are designed to construct a dual-path collaborative optimization hybrid architecture for dynamically modeling different components of the time series. Meanwhile, the model further improves an attention-guided channel feature enhancement strategy to effectively suppress redundant noise and enhance critical dimensional information. Experiments on server datasets such as SMD and PSM demonstrate that the proposed model achieves recall rates of 83.53% and 96.14%, respectively, in anomaly detection tasks, showing a significant performance improvement over mainstream baseline models.