PHYSICAL STATE MODELS OF STRUCTURAL ELEMENTS FOR PREDICTING BUILDING RELIABILITY USING GRAPH NEURAL NETWORKS

Abstract

Effective management of the life cycle of building assets requires reliable tools for predicting their technical condition. Modern approaches, in particular Building Information Modeling and the concept of digital twins, create opportunities for the application of graph neural networks. The aim of this work is to develop, through the formalization of existing theoretical provisions, validated models, and experimentally confirmed regularities, models of the physical state of structural elements of a building object as a function of time. To achieve this aim, the following tasks were accomplished: modern approaches to modeling structural degradation were analyzed in order to identify key physical laws and engineering models subject to formalization; models of the physical state were formulated for the main structural elements (columns, beams, slabs, and walls), describing the dynamics of deterioration of their characteristics over time under the influence of operational factors; and a methodology for verifying the proposed models was developed, including sensitivity analysis to key input parameters.

The results of the sensitivity analysis confirmed the physical validity of the models. It was established that for load-bearing reinforced concrete elements (columns and beams), the dominant factors are corrosion intensity and degradation of material properties. For the foundation slab, the critical factor was found to be the chemical composition of the soil, while for the wall panel, the durability of the sealant under UV radiation—which triggers a cascading failure—proved to be decisive.

A key result is the demonstrated ability of all models to clearly differentiate between two main scenarios: a “satisfactory state,” in which the structure retains a significant safety margin, and a “critical state,” in which this margin is exhausted. The practical significance of the work lies in the fact that the proposed models provide a sound computational basis for populating graph structures with physical attributes, which is a necessary prerequisite for the development and training of graph neural networks aimed at predicting the technical condition of building assets.