Temporal-Spatial Anomaly Detection in Multivariate Time Series via Image Encoding and Self-Supervised Learning

Anomaly detection in multivariate time-series data is crucial to ensuring the safety, security, and reliability of cyber-physical systems across domains such as water treatment, industrial control, and space telemetry. Despite notable progress, existing approaches often suffer from high false-positive rates, limited adaptability to complex temporal patterns, and computational inefficiencies that hinder real-time deployment. Many baseline models either fail to capture long-range dependencies or rely heavily on hand-crafted features and threshold tuning, which restricts scalability and robustness in dynamic environments. To address these limitations, this paper introduces a novel hybrid anomaly detection framework that integrates Gated Recurrent Units (GRUs), temporal convolutional layers, and an attention-guided fusion mechanism. The model is designed to extract both short- and long-term temporal features while dynamically weighting informative segments for more discriminative anomaly detection. A statistical anomaly scoring module is employed to refine detection granularity, and the overall architecture supports real-time inference with minimal latency. The framework is comprehensively evaluated on five benchmark datasets: SWaT, WADI, SMD, SMAP, and MSL. Experimental results demonstrate that the proposed approach consistently outperforms state-of-the-art models in terms of precision, recall, F1-score, and Area Under the Curve (AUC), while achieving superior inference speed. Additionally, Wilcoxon Signed-Rank tests confirm that the performance improvements are statistically significant. These results highlight the model’s effectiveness, efficiency, and suitability for real-world anomaly detection applications. This work contributes a robust, adaptive, and interpretable solution to the challenge of time-series anomaly detection in critical systems.