Anomaly Multi-classification for Microservices Based on Parallel Time Channel Attention Feedback Network

In microservice architectures, multivariate time series data constitutes a core component of system monitoring and operations maintenance, with anomaly types often exhibiting significant diversity and complexity. However, current research lacks sufficient studies on multi-class classification methods for anomalies. Therefore, this paper proposes a multi-class anomaly detection model for microservice systems based on a parallel time channel attention feedback network（PTCAFNet ）. The model uses temporal convolutional network（TCN）to extract long-term dependency features from multivariate time series, employs squeeze-and-excitation network（SENet ）to extract global features, and adopts parallel channel attention mechanisms to separately capture global and local features, and obtains global anomaly patterns and fine-grained anomaly characteristics to enhance anomaly recognition capability and robustness. Through feedback mechanisms（IM）, it strengthens feature fusion, and improves channel relationship modeling and robust feature representation. Experimental results demonstrate that PTCAFNet achieves average scores of 0.922 in Macro-F₁ and 0.963 in Micro-F₁, outperforming baseline models by approximately 36.9% and 3.3% respectively.