Anomaly Detection of Server Data Based on Depthwise Separable Convolution and Transformer

Server data is an important indicator of the server status, and its anomaly detection is greatly valuable for the normal operation of applications. However, server data is difficult to be processed effectively due to its diversity and complexity. In addition, current algorithms are difficult to find the anomalies accurately. To address this problem, we propose an anomaly detection model based on Depthwise Separable Convolution(DSC) and Transformer in this paper. The model uses sliding window partition to reorganize the data, and uses DSC and Transformer structure to learn local and global context information respectively. We reconstruct the original server data based on the learned patterns and use the reconstruction error as the final anomaly score for anomaly determination. Experiments show that our model has an F1 index of 0.92 on the public dataset SMD, proving its abilities to accurately detect the anomalies of the server data.