19:43, 27 December 2025

Artificial Intelligence to Assess Building Safety in Russia

The system classifies structural wear and reduces the risk of emergency situations.

Scientists at Perm Polytechnic University have developed a program that automatically assesses the technical condition of the exterior walls of brick buildings. The system classifies the degree of structural wear with an accuracy of up to 84 percent, the university said.

In Russia, residential buildings and other structures are regularly declared unsafe. Traditional inspection methods do not fully cope with the task, allowing hidden defects to accumulate in structures and remain unnoticed until buildings reach a critical state of deterioration.

Machine Learning Algorithms and System Architecture

The Perm Polytechnic research team created the system using artificial intelligence. To do this, they analyzed and digitized archival data from building inspections and then assembled a training dataset describing building facades across 18 parameters. Based on this analysis, the system assigns one of four condition categories defined by national standards: normal, serviceable, limited serviceability, or unsafe.

To build the intelligent system, the researchers tested five machine learning algorithms for neural networks. According to Galina Kashevarova, Doctor of Technical Sciences and professor at the Department of Structural Engineering and Computational Mechanics, the program is designed to process information in several stages.

“First, the intelligent system receives and processes initial data on the building’s condition. At the second level, it analyzes the results and calculates complex relationships between different wall condition parameters, identifying combined effects from multiple factors. The third stage generates the final assessment by assigning one of the four technical condition categories,” she explained.

Training the Model and Results

Sergey Krylov, a postgraduate researcher at the same department, added that the training process also took place in several stages.

“Initially, the system learned from 65 percent of the data we uploaded – this was the main training set. Another 20 percent was used for intermediate checks to ensure the model was not simply adapting to specific examples but identifying underlying patterns. The final stage used the remaining 15 percent of data, which the system had never seen before. To prevent overfitting – when a model memorizes examples instead of learning general rules – a neuron dropout method was applied. This approach significantly improved the system’s reliability when working with new, previously unseen buildings,” Krylov said.

On the training dataset, the model achieved an accuracy of 92.3 percent. On the validation dataset not used during training, accuracy reached 84.62 percent.

Industry and import substitution