Timing anomaly detection based on GRU-INEncoder

In the field of unsupervised timing anomaly detection, existing methods face challenges in capturing long-range dependencies and dynamic timings due to the scale of the data and multiple feature dimensions. This paper presents a novel method for timing anomaly detection that effectively extracts long-range dependencies and dynamic timing features by leveraging stacked encoders and gated recurrent units (GRUs). Moreover, it introduces a multi-branch attention mechanism to extract local and global features, thereby enhancing the model's ability to perceive information at different scales. The local attention captures fine-grained time series changes, while the global attention focuses on long-term trends and overarching patterns. Experimental results demonstrate that our method significantly outperforms existing time-series anomaly detection methods across several publicly available datasets, such as SMD, MSL, and SMAP, affirming its superiority in terms of accuracy and robustness.