Advancing battery failure diagnosis by knowledge-augmented large language models

Battery failure diagnosis is crucial for ensuring the safety and reliability of energy storage systems. However, existing approaches from electrochemical modeling to deep learning often face limitations, including heavy reliance on extensive training data, poor generalization, and interpretability issues. To address these challenges, we propose BattFailScholar, a knowledge-augmented large language model (LLM) framework for battery failure diagnosis. Our approach first constructs a case-level battery failure knowledge graph encompassing material properties, multi-source signals, and failure pathways. A knowledge-augmented generation method is then developed to enhance LLM diagnostic reasoning with failure feature-aware retrieval and optimization algorithms. Experimental results demonstrate that BattFailScholar achieves a 19.7% performance improvement in LLM-based diagnosis, with enhanced capability in alleviating long-tail problems and failure risk assessment. Moreover, the system achieves 86.2% accuracy in identifying potential failure mechanisms or causes, demonstrating strong potential for discovering failure chains and providing practical, reliable diagnostic support for battery research and development.