FCAE-Mixer: frequency-aware convolutional autoencoder with dual-branch mixing for time series anomaly detection

Abstract

The core of complex time series anomaly detection lies in the effective extraction of multi-dimensional features. However, existing methods often separate the synergistic mechanism among local temporal, global variable, and frequency domain features. To this end, this paper proposes a network that balances temporal, variable, and frequency domain features. The model innovatively combines a one-dimensional convolutional autoencoder and a multi-layer perceptron to extract local temporal transient features and global variable spatial mappings, respectively. Furthermore, an FFT-Attention module is introduced to enhance adaptive focus on low-frequency spectral modes while suppressing high-frequency noise. Finally, the dual-branch reconstruction error is used as the anomaly judgment basis, and anomaly samples are identified via the reconstruction error, which jointly benefits from both time-domain morphological and frequency-domain spectral processing. Evaluation results on seven benchmark datasets demonstrate that the proposed method achieves an average F1 score (without point-adjustment) of 0.3317, outperforming all compared baseline methods with a 10.50% relative F1 improvement over the suboptimal baseline Transformer with an average F1 of 0.3002, demonstrating the effectiveness of multi-dimensional joint modeling in time series anomaly detection.