ReDTF-AD: Reconstruction-Based Decomposition and Time–Frequency Fusion for Time Series Anomaly Detection

Time series anomaly detection aims to identify deviations from the normal distribution of temporal data. Reconstruction error is a natural and practical anomaly criterion, and using reconstruction error as an anomaly criterion is a well-established and practical paradigm. However, existing reconstruction-based methods often fail to capture complex structures in high-dimensional time series data and typically lack in-depth analysis of periodicity, limiting detection accuracy. To address these challenges, we propose ReDTF-AD, a novel approach that integrates reconstruction error with time–frequency fusion. Specifically, the input series is decomposed into seasonal and trend components. For seasonal components, we designed a time–frequency fusion block (TFFB) to enhance frequency-domain features while preserving residual time-domain information, ultimately achieving the fusion of time–frequency information. The Top-k transformation converts 1D sequences into 2D representations based on the periodicity of time series data, enabling deeper analysis of intra-period and inter-week variations through our newly proposed Split Concat Block (SCBlock). For the trend term, a linear module captures long-term patterns. In the unsupervised time series anomaly detection experiments based on reconstruction error, ReDTF-AD shows competitive performance on three public benchmark datasets.