DCFAD: Dual-Path Cross-Attention with Learnable Fractional Fourier Transform for Time Series Anomaly Detection

Abstract

Multivariate time series anomaly detection is crucial for failure diagnosis, network security, and intelligent operations. Reconstruction-based approaches are well suited for unsupervised scenarios, as they detect anomalies without labeled data and model complex temporal patterns through reconstruction. However, despite recent progress in point-wise anomaly detection and frequency-enhanced representations, existing methods largely rely on fixed representation domains, which may limit their ability to separate diverse pattern-wise anomalies from normal patterns and produce robust anomaly scores. We propose DCFAD, a unified framework that integrates a learnable fractional Fourier transform (FrFT) and a dual-path cross-domain attention mechanism. DCFAD learns anomaly-sensitive intermediate representations between time and frequency domains by adaptively adjusting the fractional order of FrFT during training. Meanwhile, the dual-path cross-domain attention establishes cross-domain correspondences from temporal and variable dimensions, enabling selective fusion of time-domain and fractional-domain features for reliable anomaly reconstruction. Extensive experiments on 10 real-world datasets with multiple evaluation metrics demonstrate the state-of-the-art overall performance of DCFAD in multivariate time series anomaly detection. Notably, DCFAD achieves a 9.5 percentage-point improvement in average AUC-ROC over the strongest AUC-ROC baseline while maintaining efficient inference. The complete code has been released at https://github.com/Jia-xw1/DCFAD.