TCN-based Lightweight Log Anomaly Detection in Cloud-edge Collaborative Environment
Jining Chen, Weitu Chong, Siyu Yu, Zhun Xu, Chaohong Tan, Ningjiang Chen · 2022
In the cloud-edge collaboration system, logs help developers diagnose system failures, and make abnormal diagnosis decisions as soon as possible for edge devices with a low tolerance for failure recovery time. As the computing power and storage of the edge devices are limited, the existing log-based anomaly detection approaches are difficult to do well in model size, computing power requirements, and accuracy at the same time. For example, the approach based on PCA has low anomaly detection accuracy, while Transformer-based models have huge parameters. In this paper, we propose a lightweight log-based anomaly detection model based on a compressed Temporal Convolutional Network (TCN), called EdgeLog. Edgelog replaces the fully connected layer of TCN with a Global Average Pool(GAP), and uses a sparsification strategy to compress the TCN network. In addition, a tree structure-based anomaly detection model deployment strategy for the cloud-edge collaborative system is proposed. Experimental results on 4 public datasets demonstrate the effectiveness and efficiency of Edgelog. Compared with the baseline, Edgelog’s model size is reduced by around 75%, FLOPS is reduced by around 97%, and it surpasses the baseline in average Precision, F-value, and Recall.