Decision Making in Structural Health Monitoring Using Large Language Models

Abstract

Structural health monitoring (SHM) is fundamental in decision making for damage identification and prescriptive maintenance of civil infrastructure. Traditional decisionmaking methods often fall short in integrating heterogeneous sensor data, inherent to SHM, and the natural language processing required to interpret, e.g., inspection reports, expert assessments, and historical documentation relevant to prescriptive maintenance. To address these limitations, this study introduces a framework for integrating large language models (LLMs) into SHM workflows, facilitating decision making in damage identification and prescriptive maintenance. The proposed framework links convolutional neural networks (CNNs) with generative LLMs. Unlike approaches based solely on prompt engineering, this study applies task-specific fine-tuning via low-rank adaptation (LoRA) to the Mistral-7B-Instruct-v0.1 model, using CNN-generated output and damage metadata as input. The results demonstrate – apart from the proof of concept – a successful generalization to previously unseen CNN-generated output, enabling context-sensitive damage identification. In conclusion, integrating LLMs into SHM workflows allows synthesizing heterogeneous sensor data and natural language, thus enhancing decision making for damage identification and prescriptive maintenance of civil infrastructure.