Leveraging Generative AI for Predictive Maintenance - Building a Knowledge Base for Fault Diagnosis

Abstract

Fault diagnosis is a complex challenge for industrial production. This thesis develops and evaluates a predictive maintenance assistant integrating large language models (LLM) with retrieved-generation techniques (RAG). By constructing a unified knowledge base comprising sensor data, event logs, and equipment manuals, the system enhances fault diagnosis in industrial settings. The system analyzes sensor data and links it to event logs, matching sensor data with faults. Meanwhile, it connects faults with equipment manuals via RAG, forming a unified knowledge framework. It generates readable and accurate fault diagnostics via LLMs and searching relevant technical documents. It is adaptive, transferable, and capable of integrating specific knowledge. Experiments conducted using a simulated drone assembly production line demonstrate significant improvements in diagnostic accuracy, interpretability, and reliability, effectively addressing common issues such as hallucinations and unsupported claims found in traditional LLM applications. The findings highlight the practical feasibility of deploying advanced AI-driven predictive maintenance solutions, emphasizing the importance of semantic richness and structured knowledge integration.