Automation of monitoring construction and installation works using neural network image processing and cloud–edge architecture for real-time updating of schedules and cost estimates

Abstract

The study examines ways to improve the timeliness and reliability of monitoring construction and installation works in large-scale projects. The aim is to develop and evaluate an integrated control system that automatically records actual volumes and immediately reflects them in managerial documentation. The methodological basis relies on the concept of a digital twin of the construction process and the use of neural network image analysis combined with a cloud-edge architecture. The materials consist of field test data from twelve residential and commercial sites, including more than three hundred thousand annotated photographs obtained from unmanned aerial vehicles and wearable cameras. The core of the system is a convolutional neural network based on the YOLOv5 architecture, trained to recognize reinforcement, formwork installation, concreting, masonry of enclosing structures, installation of windows, and façade systems. Image processing is performed on edge devices with graphical accelerators, which generate compact structured data and transmit them to the cloud platform. The cloud environment implements modules for matching recognition results with the building information model, automatically updating the schedule network diagram, cost estimates, and acceptance certificates of completed works. Recognition quality was evaluated according to standard machine learning metrics and achieved an average F1 Score of about ninety-six percent, maintaining acceptable accuracy even under challenging shooting conditions. Comparison with traditional control demonstrated a reduction of the full cycle from data collection to updating schedules and estimates by more than twenty-two times, from approximately 11 hours to less than 30 minutes. The greatest effect is achieved at the stages of volume identification and documentation updating due to the transfer of inference to the edge and integration with cloud project management services. Economic analysis using a typical project showed a reduction of more than two-thirds in direct costs for control procedures and defect correction, as well as a sharp decrease in penalty charges for schedule delays. The total estimated savings amount to approximately fifteen million rubles, corresponding to a reduction in the total cost of construction and installation works by about twelve to fifteen percent. The paper discusses application limitations related to image degradation during precipitation and fog, incomplete coverage of internal engineering systems, and the need to expand the training dataset. The prospects of integrating the system with predictive analytics and robotic tools are noted, which should ensure a transition to a closed-loop control and a fully data-driven construction process.