Automated Feature Extraction fromReinforcement Drawings for Machine Learning-Based Cost Estimation

Sammanfattning

Accurate price estimation for engineer-to-order steel products is currently performed manually by experienced estimators working from structural drawings, a process that is time-consuming and varies across personnel. This thesis develops an automated pipeline that extracts cost-relevant features directly from Tekla Structures reinforcement drawings in PDF format and uses them to train regression models for welding cost prediction. Because Tekla structures render all drawing content as vector graphics rather than searchable text, features are recovered through a combination of Canny edge detection, the Probabilistic Hough Transform, and DBSCAN clustering. Seventeen features, including steel mass, bill-of-materials complexity derived from ERP records, and weld intersection count extracted from image analysis, are linked to historical cost records from 367 orders across three production facilities. XGBoost and OLS regression are trained andevaluated using MAE, RMSE, and MAPE with five-fold cross-validation. XGBoost achieves a cross-validation MAE of 13,793 SEK and a test-set MAPE of 110.4%, with OLS performing comparably at 15,406 SEK.