CICFormer: An efficient unsupervised performance anomaly detection model based on temporal context attention and interactive convolution for fluctuating cloud environments

Abstract

Anomaly detection in real-time monitoring metrics is crucial for ensuring the stability and service quality of cloud computing systems. However, owing to the dynamic nature of cloud computing resources, interdependencies among high-dimensional monitoring indicator data, and the sparsity of anomalies, anomaly detection tasks face significant challenges. Traditional methods often rely on static features or single time-scale analyses, failing to fully consider the complex contextual information and temporal dependencies between monitoring indicator data, making it difficult to accurately identify potential anomalies in cloud monitoring metrics. To address this issue, this paper proposes an efficient unsupervised performance anomaly detection model based on temporal context attention and interactive convolution transformer (CICFormer) for fluctuating cloud environments. This model adopts an unsupervised temporal anomaly detection method that integrates the temporal context-aware mechanism (TCAM), squeeze-and-excitation (SE), and interactive convolutional block (ICB). TCAM leverages temporal pooling and channel attention to enhance understanding of key time points, SE improves sensitivity to relevant variables, and ICB captures both local and global features using multi-scale convolutions. Extensive experiments on five public datasets show that CICFormer outperforms 18 baseline methods, achieving a 4.26 % improvement in the average F1 score. These results demonstrate that CICFormer is highly effective for real-time anomaly detection in cloud environments.