SEA-FGT: Frequency-Guided Transformer with Semantic Expert Augment for Time Series Anomaly Detection

Multivariate Time Series (MTS) anomaly detection is fundamental to industrial monitoring and intelligent operations. However, practical deployment remains difficult, as real-world MTS exhibit complex inter-channel dependencies and anomaly-induced perturbations, which are often subtle in time domain. Moreover, the implicit semantics and heterogeneity of time series make it challenging for empirically designed augmentations to increase pattern diversity while preserving semantic consistency. To address these challenges, we propose SEA-FGT, a Frequency-Guided Transformer with Semantic Expert Augmentation for MTS anomaly detection. Specifically SEA-FGT incorporates frequency-aware modeling to capture stable cross-channel dependencies and anomaly-related spectral perturbations, while introducing a semantic expert augmentation mechanism to adaptively align augmentations with diverse temporal semantics. In addition, we provide the theoretical analysis for the semantic expert augmentation mechanism from an information-theoretic perspective, demonstrating that it extends diversity criteria into conditional formulations and proving a bound on representation robustness of fidelity. Extensive experiments on multiple real-world datasets demonstrate that SEA-FGT has competitive performance against state-of-the-art baselines across diverse evaluation metrics, while additional explorations and visualizations further corroborate its effectiveness.