Accurate and Interpretable Log-Based Fault Diagnosis Using Large Language Models

Log-based fault diagnosis is essential for ensuring system reliability and resilience. However, current methods only provide fault diagnosis results without explanations, which undermines their credibility. Large language models (LLMs) have extensive pre-trained knowledge and show potential in log analysis, yet they cannot be directly applied to log-based fault diagnosis due to limited specialized capabilities and domain-specific insights. Furthermore, LLMs have limitations in context length and are too diverse to select a suitable one. To address these issues, this paper presents <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LogInsight</i>, a framework that enables accurate and interpretable log-based fault diagnosis using LLMs. We fine-tune a medium-sized, open-source LLM to incorporate domain expertise and leverage its interpretive capability. Additionally, we design a Fault-Oriented Log Summary (FOLS) module to extract essential information from log sequences, mitigating LLMs’ context length limitation. Extensive evaluations on two public datasets and a real-world production dataset demonstrate <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">LogInsight</i>’s superiority over state-of-the-art methods in both performance and interpretability.