Research on Anomaly Detection Based on System Log Data

As modern software scales up and becomes increasingly complex, using log files to detect anomalies and ensure software reliability has become a research hotspot. However, the accuracy of existing log anomaly detection algorithms may decrease with system updates. In this paper, we propose a log anomaly detection model called LogCGA. First, we use the Drain algorithm to extract log templates, then divide log events into log sequences using fixed and sliding windows. We then extract statistical features and semantic features from the log using template counters and FastText language models, respectively. Finally, we input the feature vector into a parallel CB model for detection. We conduct experiments on public datasets and show that LogCGA significantly improves the accuracy and stability of anomaly detection compared to benchmark models.