From Prediction to Prescription: Large Language Model Agent for Context-Aware Maintenance Decision Support

Predictive analytics with machine learning approaches has widely penetrated and shown great success in system health management over the decade. However, how to convert the prediction to an actionable plan for maintenance is still far from mature. This study investigates how to narrow the gapbetween predictive outcomes and prescriptive descriptions for system maintenance using an agentic approach based on the large language model (LLM). Additionally, with the retrieval-augmented generation (RAG) technique and tool usage capability, the LLM can be context-aware when making decisions in maintenance strategy proposals considering predictions from machine learning. In this way, the proposed method can push forward the boundary of current machine-learning methods from a predictor to an advisor for decision-making workload offload. For verification, a case study on linear actuator fault diagnosis is conducted with the GPT-4 model. The result demonstrates that the proposed method can perform fault detection without extra training or fine-tuning with comparable performance to baseline methods and deliver more informatic diagnosis analysis and suggestions. This research can shed light on the application of large language models in the construction of versatile and flexible artificial intelligence agents for maintenance tasks.