open access

https://hdl.handle.net/2115/100475

Integrating Point Cloud Features with Large Language Models for Bridge Understanding

英語

bridge engineering

2026-07-01

Proceedings of the 10th International Symposium on Reliability Engineering and Risk Management (ISRERM 2026)

Automated structural inspection of bridges using LiDAR point clouds holds great promise for reducing inspection costs and improving safety. However, existing point cloud understanding methods are designed for general-purpose object-centric or indoor scenes and fail to exploit the rich structural priors inherent in civil infrastructure. We present BridgeLLM, the first multimodal large language model (LLM) framework specifically designed for bridge structural point cloud understanding. The core of BridgeLLM is BridgeSPT (Structure-Preserving Transformer), a novel hierarchical encoder that integrates five bridge-domain inductive biases: (i) adaptive hash-based voxel tokenization, (ii) parameter-free PCA structural geometric descriptors (linearity, planarity, scattering), (iii) additive geometric conditioning at every Transformer stage, (iv) dual Morton-curve serialization sensitive to bridge elongation, and (v) a learnable query bottleneck for compact LLM-aligned representations. Training follows a two-stage protocol coupling feature alignment with LoRA-based instruction tuning on Qwen2.5-7B-Instruct. Evaluated on ten real-world LiDAR bridge scans with 10,000 instruction-response pairs, BridgeSPT achieves the lowest inference latency while generating accurate structural descriptions across diverse question types-bridge classification, component identification, dimensional estimation, and free-form inspection reporting.

会議発表論文

Version of Record