Unsupervised anomaly detection in multivariate time series is critical for the reliability of the Industrial Internet of Things (IIoT) and IT operations. However, existing reconstruction-based approaches struggle with anomaly contamination, complex dependency modeling, and non-stationary interference. To address these challenges, this paper proposes TFD-CAD, a novel Collaborative Time-Frequency Dual-Branch Anomaly Detection framework. Distinct from methods that process domains in isolation, TFD-CAD introduces a collaborative masking mechanism where temporal features explicitly guide high-frequency masking to purify representations against contamination. To capture intricate dependencies, a multi-scale fusion module is designed within the frequency branch to model intra-band details and inter-band harmonic correlations simultaneously. Furthermore, non-stationarity is addressed via a residual-focused reconstruction strategy. By leveraging time-series decomposition and static memory-guided attention entropy regularization, the model effectively disentangles legitimate macro-trends from anomalies. Extensive experiments on multiple real-world benchmark datasets demonstrate that TFD-CAD achieves state-of-the-art performance, significantly outperforming existing baselines in terms of detection accuracy and robustness.
Chen et al. (Wed,) studied this question.