Evaluation of tunnel rock mass integrity using multi-modal data and Generative Large Language Model: TunnelRIP-GPT

Abstract

The prediction of tunnel rock mass integrity is critical for ensuring construction safety. Tunnel rock mass offers a vast amount of multimodal data, yet accurately integrating these multimodal to predict rock mass integrity remains challenging. This study proposes a novel framework combining an Enhanced Computer Vision (CV) model and interactive Generative Large Language Models (LLMs) to address this gap. First, the Stable Diffusion method is employed to augment fracture image datasets, mitigating data imbalance issues. An improved CV method achieves 83% test accuracy in rock fracture segmentation. Subsequently, seven parameter sets-encompassing apparent, internal, physico-mechanical, and design characteristics are constructed from geological sketches, ground-penetrating radar, and drilling reports. A Tunnel rock integrity prediction GPT method is developed, which achieves a higher prediction accuracy of 92% and overcomes the limitation of conventional Transformer and CNN methods, which cannot directly integrate these multi-modal data (including images, numerical values, and texts). Finally, generalizability is confirmed through validation on untrained datasets. The study further evaluates the predictive efficacy of individual parameter sets, enabling tailored solutions for diverse tunnel conditions, and demonstrates the potential of generative LLMs to replace traditional CV and digital image processing in geotechnical applications.