Anomaly Detection in Incomplete BIM Models Based on Contrastive Learning

Abstract

Building Information Modeling (BIM) has become a cornerstone in digital construction, yet the prevalence of incomplete data poses significant challenges for reliable anomaly detection and quality assurance. This study aims to develop a robust anomaly detection framework tailored for BIM environments characterized by missing, inconsistent, or heterogeneous information. A deep contrastive learning architecture is proposed, integrating graph-based representation learning with explicit missingness modeling and data augmentation strategies. The model constructs positive and negative pairs from augmented BIM samples, enabling the network to learn invariant and discriminative features despite varying degrees of data incompleteness. Experimental evaluation is conducted on a comprehensive BIM dataset comprising both authentic and systematically simulated missing data. Results indicate that the proposed framework achieves superior precision, recall, F1-score, and AUC compared to traditional machine learning and deep learning baselines, particularly under conditions of high or structured missingness. Ablation studies demonstrate the critical role of each architectural component, and robustness analysis confirms the method’s stability when facing extreme data loss. These findings underscore the practical value of the approach for real-world BIM quality control, as it enables accurate and reliable detection of semantic and topological anomalies without requiring full data integrity. The research concludes by outlining prospects for deploying the framework at scale, integrating with industrial BIM platforms, and addressing more complex anomaly types in future work.