Time Series Anomaly Detection Based on Entropy-Sparsified Time-Frequency Fusion and MsRwGWO Meta-Optimization

Addressing the core challenges in multivariate time series anomaly detection within complex industrial environments, such as redundant time-frequency feature fusion, significant noise interference, and difficulties in model hyperparameter tuning, this study proposes a detection framework (TFUL) based on entropy-sparsified time-frequency fusion and a Multi-strategy Random Weighted Grey Wolf Optimizer (MsRwGWO). The main contributions of this work include: 1) A dual-domain entropy sparsification fusion mechanism is designed, which dynamically evaluates and filters crucial temporal segments and frequency components via information entropy, enabling adaptive and redundancy-resistant feature fusion. 2) A heterogeneously collaborative feature extraction network is constructed. The temporal branch, SoftShapeNet, integrates multi-scale convolutions and a Mixture of Experts (MoE) to capture local polymorphic shapes, while the frequency branch, FrequencyDomainProcessor, employs a learnable Mahalanobis distance to model nonlinear spectral dependencies among channels, surpassing the limitations of fixed transformations. 3) The MsRwGWO meta-optimization strategy is proposed, which incorporates dynamic weighting and multi-strategy perturbation mechanisms, significantly enhancing the efficiency and quality of hyperparameter search. Experiments conducted on several public datasets demonstrate that the pro-posed method outperforms mainstream comparative models in terms of detection accuracy and robustness, providing an effective solution for industrial time series anomaly detection.