Machine Learning Methods for Organizational and Technological Design of Housing Stock Major Repair

This study develops a methodology for creating customized organizational and technological solutions for capital repairs of structural elements in residential buildings within large-scale regional renovation programs. The research addresses the critical need to transition from standardized approaches to condition-based solutions that account for the actual technical state of each structural element through comprehensive defect analysis. The proposed methodology fundamentally differs from conventional approaches by utilizing real structural conditions and defect patterns as primary input data rather than standard technical specifications. This innovative approach requires novel analytical frameworks, as identical structural components may exhibit different defect profiles necessitating distinct repair strategies. Through systematic comparison of machine learning techniques, we developed a hybrid SOM-Random Forest model that combines the pattern recognition capabilities of Kohonen self-organizing maps with the predictive accuracy of random forest algorithms. The implemented solution enables automated clustering of objects with similar defect characteristics and optimal repair strategy selection. The methodology was validated using a synthetic dataset of 61 flat roof structures, resulting in the identification of five distinct repair clusters. Additional testing on a rolled roof case study confirmed the approach’s effectiveness, demonstrating accurate prediction of required repair measures. The results show significant improvement in renovation planning precision while maintaining the scalability needed for regional implementation programs.