@article{Zheng2026, 
author = {Zeyu Zheng and Jie Lu and Fengtai He and Bingyi Wang and Zhen Yu and Liang Ji and Yang Zhao},
title = {Confidence-driven HVAC component recognition integrating LLMs and computer vision},
year = {2026},
journal = {Building Simulation},
volume = {19},
number = {6},
pages = {1433-1453},
keywords = {HVAC drawings, component recognition, large language model, computer vision, confidence-driven reasoning},
url = {https://www.sciopen.com/article/10.1007/s12273-026-1425-0},
doi = {10.1007/s12273-026-1425-0},
abstract = {Accurate recognition of components in heating, ventilation, and air conditioning (HVAC) drawings is crucial for building information digitization. Existing recognition methods lack the ability to integrate multi-source information, such as component symbols, textual annotations, connection relationships, and domain knowledge, into a unified reasoning process. As a result, they exhibit poor performance when applied to HVAC drawings with diverse and non-standardized component symbols. To address these challenges, this paper proposes a confidence-driven method integrating large language models (LLMs) with computer vision (CV). A structured representation approach is designed to preserve spatial and semantic relationships in HVAC drawings, enabling the LLM to reason components beyond raw image inputs. A confidence-weighted mechanism assigns confidence scores to multi-source information, allowing high-confidence evidence to iteratively guide the refinement of uncertain component classifications. A dynamic knowledge infusion strategy is introduced to integrate HVAC-specific knowledge into the reasoning loop, allowing the reasoning process to produce domain-consistent component recognition. The proposed method is compared to conventional CV models on open-access HVAC drawing datasets. Results show that the true positive rate for every component category exceeds 0.89. Significant improvements of 46.4% in recall, 14.5% in precision, and 36.0% in F1-score over the baseline. Additionally, the model achieves a macro-average area under the curve (AUC) of 0.977, verifying its robust discrimination capability. The proposed framework effectively enhances the accuracy and robustness of HVAC component recognition, providing a scalable solution for automated building digitization.}
}