Improvement of multi-parameter anomaly detection method: Addition of a relational token between parameters
Hironori Uchida, Keitaro Tominaga, Hideki Itai, Yujie Li, Yoshihisa Nakatoh · Cognitive Robotics · 2025
In the continuous development of systems, the increasing volume and complexity of data that engineers need to analyze have become significant challenges. As a solution, a considerable amount of research has been conducted on anomaly detection in logs through automated methods. However, due to the limited variety of datasets, most research has focused on sequence anomalies in logs, with little attention paid to anomaly detection based on parameters within logs. To address this issue, we prepared a labeled dataset specifically for parameter-based anomaly detection and propose a novel method utilizing BERTMaskedLM. Notably, since it is challenging to label continuously changing logs in system development, we propose a method that can learn without labeled data. Previous studies have used BERTMaskedLM to learn relationships between parameters in logs with multiple parameters for anomaly detection. However, a known issue is that when the ranges of numerical parameters overlap, detection accuracy decreases. In our experiment, by adding tokens that encode the relationships between parameters, we improved the independence of parameter combinations and enhanced the accuracy of anomaly detection (increasing the F1-Score by more than 0.002). Additionally, we visualized the influence of the added tokens and conducted experiments using a new dataset to assess the reliability of the proposed method.