AI Agents Diagnose Bridge Damage Using QLoRA and Causal Knowledge

Diagnosing the root causes of bridge deterioration, such as salt intrusion or fatigue cracking, often requires years of expert tacit knowledge because these causes are not directly visible. This research aims to automate this complex causal reasoning by developing a Damage Cause Encoder. This encoder is designed to classify ten different damage causes based solely on visible damage descriptions, making it suitable for autonomous bridge diagnostic agents. The proposed approach integrates three key components. First, a large language model extracts causal knowledge triples (e.g., damage caused by cause) from numerous diagnostic manuals and stores them in a vector database. Second, during both training and inference, relevant causal triples are retrieved and combined with the visible damage descriptions, effectively converting implicit domain knowledge into explicit context for the encoder. Finally, the researchers conducted a rigorous comparison of fine-tuning methods, including LoRA, QLoRA, and QA-LoRA. QLoRA emerged as the optimal choice, achieving identical test accuracy to full-precision LoRA while offering significant advantages: 11% faster inference and 72% lower GPU memory consumption. This efficiency allows for high-accuracy diagnostic agents to be deployed on consumer-grade hardware at the edge, making advanced bridge diagnostics more accessible.