OTOMASI PENYARINGAN IRRELEVANT CLASHES PADA KOORDINASI BIM STRUKTUR DAN MEP MENGGUNAKAN METODE HIBRIDA RULE-BASED REASONING DAN MACHINE LEARNING

Azkianof Royhaan Ramadhan, Dr. Toriq Arif Ghuzdewan, ST, M.Sc.E.

2026 | Skripsi | TEKNIK SIPIL

Proses deteksi benturan pada perangkat lunak BIM sering kali menghasilkan ribuan peringatan, di mana mayoritasnya merupakan benturan palsu atau tidak relevan (irrelevant clashes). Peninjauan manual terhadap clash ini menguras ratusan jam kerja praktisi dan memicu kelelahan. Permasalahan utama yang diangkat adalah bagaimana cara menyaring benturan palsu secara otomatis dan akurat dari laporan XML bawaan tanpa menimbulkan kebocoran data (data leakage) akibat penggunaan atribut koordinat spasial absolut. Penelitian ini bertujuan mengembangkan model kecerdasan buatan berbasis pendekatan hibrida untuk mengklasifikasi tingkat relevansi benturan dan meminimalisasi beban kerja koordinator BIM. Metode penelitian mengintegrasikan Rule-Based Reasoning (RBR) sebagai pemberi aturan awal, dan Machine learning (XGBoost) sebagai algoritma pembelajar utama. Dataset pemodelan terdiri atas 904 sampel benturan berlabel hasil ekstraksi XML Autodesk Navisworks. Model secara khusus menghindari ekstraksi dimensi volume (bounding box) dan mengandalkan rekayasa fitur relasional, seperti interaksi pasangan elemen dan kepadatan persilangan (clash hub). Teknik analisa data berfokus pada penanganan kelas tidak seimbang (class imbalance) melalui fungsi Cost-Sensitive Learning di dalam XGBoost tanpa memanipulasi kepalsuan data sintetik, yang dievaluasi menggunakan Confusion Matrix dan analisis Feature Importance. Hasil pengujian menunjukkan batas keputusan klasifikasi dengan perolehan Accuracy 0.9227, Precision 0.8841, Recall  F1-Score 0.8971, dan ROC-AUC  0.9753. Secara kuantitatif, implementasi pengujian model pada dataset secara utuh sukses mengeliminasi 61.9% peringatan irrelevant clash secara otonom, sekaligus mengamankan peringatan pada benturan struktural kritis. Analisis mengungkapkan pengelompokan hierarki disiplin utilitas MEP merupakan prediktor paling determinan. Kesimpulannya, integrasi RBR-XGBoost terbukti amat efektif mereduksi beban kerja peninjauan manual secara masif dan menawarkan solusi otonomasi koordinasi desain yang jauh lebih efisien.

The clash detection process in BIM software often generates thousands of warnings, the majority of which are false or irrelevant clashes. Manual review of these clashes consumes hundreds of engineering hours and induces fatigue. The main problem addressed is how to automatically and accurately filter out false clashes from native XML reports without causing data leakage due to the use of absolute spatial coordinate attributes. This study aims to develop an artificial intelligence model based on a hybrid approach to classify clash relevance levels and minimize the workload of BIM coordinators. The research methodology integrates Rule-Based Reasoning (RBR) to provide initial rules, and Machine Learning (XGBoost) as the primary learning algorithm. The modeling dataset consists of 904 labeled clash samples extracted from Autodesk Navisworks XML. The model specifically avoids extracting volume dimensions (bounding boxes) and relies on relational feature engineering, such as element pair interactions and intersection density (clash hubs). The data analysis technique focuses on handling class imbalance through the Cost-Sensitive Learning function within XGBoost without manipulating synthetic data, evaluated using the Confusion Matrix and Feature Importance analysis. The testing results demonstrate a classification decision boundary with an Accuracy of 0.9227, Precision of 0.8841, F1-Score of 0.8971, and ROC-AUC of 0.9753. Quantitatively, the implementation of model testing on dataset successfully eliminated 61.9% of irrelevant clash warnings autonomously, while simultaneously securing warnings for critical structural clashes. The analysis reveals that the hierarchical grouping of MEP utility disciplines is the most determinant predictor. In conclusion, the integration of RBR-XGBoost is proven highly effective in massively reducing the manual review workload and offers a much more efficient automated design coordination solution.

Kata Kunci : Building Information Modeling (BIM), Clash Detection, Machine Learning, XGBoost, Rule-Based Reasoning (RBR)