Research is aimed at identifying spatial segmentation, classification and geoinformation analysis of defects in cement concrete coatings with a wear ball. The technique is intended to serve as the basis for the formation of a digital model that will become a tool for predictive diagnostics. The core of the method contains the YOLOv8-seg neural network, optimized for instance segmentation tasks in real time. A new technological package has been developed, which includes: 1) mobile recording with synchronous GNSS binding; 2) forming a specialized dataset with five classes, adapted to the characteristics of the combined coating; 3) active modeling on marked polygonal masks; 4) post-processing and combining masks from different frames; 5) generation of vector balls of defects in GeoJSON format and their integration into GIS; 6) automatic breakdown of key indicators of the technical state (implosion index, number of defects). The testing was carried out on the plots of the highway M-07 Kiev – Kovel. The system provided pixel segmentation of defects with an average accuracy (mAP50-seg) of 0.87 on the test sample. A quick analysis revealed that more than 58.2 % of the surface area is found in the reference station, while partially constructed plots become 32.6 %, and completely constructed plots – 9.2 %. A deep spatial analysis revealed a statistically significant correlation between zones of concentration of defects of the “designed surface” type and points of dynamic stress (knuckles, joints, plating areas). This made it possible to classify 92 % of all significant ruins as «false lands». The developed methodology has brought the possibility of complete automation to the process of extracting objective, metrologically valid data about the coating mill. The results of the investigation can only be used to clearly describe the country, and also, through extensive analysis, indicate the reasons for it. This allows you to move from routine repairs to complete, technologically advanced repairs. The system forms a digital trace of the object, which is a critically important first step until the creation of a predictive digital model. Further research will focus on the integration of visual inspection data with the results of non-contact flaw detection methods (for example, ground penetrating radar probing) for assessing not only the surface, but also the internal balls of the structure
Oksen et al. (Tue,) studied this question.