Abstract: Existing deep learning-based time series anomaly detection methods exhibit limitations in both the collaborative mode-ling of local and global features,as well as in computational efficiency for long sequences.To address these issues,this paper proposes a hybrid-architecture anomaly detection model.Firstly,for multi-scale feature extraction,depthwise separable convolution is employed to capture local detail features,combined with adaptive pooling to enhance the response capability to burst anomalies.Secondly,a bidirectional temporal modeling mechanism is designed,which integrates bidirectional context features through Bi-LSTM and utilizes gated Dropout to mitigate the risk of overfitting during long-sequence training.Furthermore,a hierarchical sparse global attention module is designed,leveraging local window attention to capture time-frequency features,while employing multi-head mechanisms and residual connections to optimize gradient propagation stability.Finally,a dynamic feature fusion method is proposed and integrated with a chunk-based processing framework to achieve collaborative optimization of detection accuracy and computational efficiency.Experimental results on four public time-series datasets demonstrate that the proposed model achieves significant improvements across various performance metrics compared to existing methods,while also exhibiting strong robustness and generalization capability.

Key words: Time series anomaly detection, Hybrid neural network, Multi-scale feature fusion, Layered sparse attention