A Deep Learning-based Auditing System for BIM Components Library

The BIM Component Library serves as a collaborative platform essential for sharing BIM resources. However, the platform's integrity is often compromised by user-uploaded components that contain errors such as misclassification, mismatched thumbnails, and inconsistent screenshots. Traditional correction methods rely on manual review, which is time-consuming, inefficient, and plagued by prolonged auditing processes. To address these challenges, this study introduces an automated auditing system for BIM components, leveraging deep learning techniques. The system employs advanced deep learning algorithms to analyze and extract image features of BIM components, creating a vector database coupled with a post-processing rules engine. This innovative combination enables the system to autonomously identify category discrepancies and ensure image uniformity. It harnesses the power of AI algorithms to facilitate automatic verification of component information. The system enhances its capability through the adoption of various pre-trained models, achieving effective zero-shot feature extraction for BIM components. Further refinement is achieved through transfer learning, which fine-tunes the model to enhance the quality of image embeddings. As a result, this approach significantly reduces audit-related costs and dramatically improves the efficiency of the auditing process.