Automation of the construction control process using artificial intelligence and augmented/virtual reality technologies

Introduction. Automation of construction control is an important task for capital construction, the oil and gas industry, and mechanical engineering. The use of augmented (AR) and mixed (MR) reality technologies allows for the visual comparison of BIM models with actual objects on the construction site, but existing systems face limitations in terms of positioning accuracy (deviations of up to 71 mm, compared to the required 15 mm), loss of reference during movement, and the inability to accurately assess geometric inconsistencies on complex surfaces. Purpose of the work. Development and formalization of a loss function that contains information about the geometric features of an object, and positioning algorithms for construction control tasks using AR/MR technologies that take into account the internal geometry of the surface: geodesic distances, consistency of normals, local curvature, and spectral characteristics of the Laplace–Beltrami operator. Methods. A hybrid approach is proposed, combining voxel maps of the environment and a surface representation in the form of a triangular mesh. The final loss function for training the positioning system is a weighted sum of five components: Euclidean distance, geodesic component, normal component, curvature, and spectral component. The model is trained using gradient descent. Results. The analysis showed that the use of only Euclidean metrics does not allow to correctly estimate the deviations between the TIM model and the real object, especially for surfaces with complex geometry, high curvature and sharp boundaries. The proposed combination of geometric functionals provides a more informative signal for positioning algorithms, reducing the mean-square error of alignment, and also allows to preserve engineering-significant details that are smoothed or disappear in the traditional approach. Conclusions and prospects. Geometrically-informed methods eliminate a number of fundamental drawbacks of AR/MR construction control systems: loss of positioning during movement, insufficient accuracy on complex surfaces, and ignoring the internal geometry of objects. A promising area is the development of hybrid architectures and the integration of differential geometry methods into simultaneous localization and mapping algorithms.