Priority 2030: Moscow Polytechnic University Scientists Train Neural Network to Detect Dangerous Cracks in Metal
A new intelligent system combining a neural network with fuzzy logic algorithms promises to automatically identify cracks and other defects in cast metal parts directly on the production line - with greater accuracy and speed than human inspectors. The system is being developed by Sergey Kuzovov, an Associate Professor at Moscow Polytechnic University. The project was recently featured by Gazeta.ru..
In mechanical engineering, quality control of castings still relies heavily on manual inspection. A technician examines a part, assesses its surface, and either sends it on for further processing or rejects it. This method is inherently unreliable: eyes tire, lighting conditions change, and cracks on complex surface textures can easily go unnoticed. A hot crack in a cast component is not merely a cosmetic flaw. In industries like aviation, automotive, and energy, such a defect missed during quality control can lead to serious accidents once the part is in service. Furthermore, production volumes at modern plants make it virtually impossible to inspect every single part manually with the required level of thoroughness.
Standard computer vision algorithms have proven only partially effective for this task. While they perform reasonably well under controlled conditions, they struggle when variables become more complex - when the surface is oxidized, the material is non-uniform, the defect's boundaries are blurred, or a crack mimics the part's natural surface relief. This leads to either false positives or missed defects - both of which are costly for manufacturers.
"We are combining a convolutional neural network, which analyzes the part's image, with fuzzy logic, which is designed to handle uncertainty," explained Sergey Kuzovov. "The system doesn't just detect a crack; it assesses the degree of danger based on context - considering surface characteristics, oxidation levels, and material type. This represents a fundamentally different level of diagnostics compared to current methods."
The project is based on deep learning architectures that have demonstrated significant effectiveness in image analysis. The neural network will be trained on a labeled database of images showing defective parts. Researchers are collecting and annotating these images, recording the location, shape, and nature of each defect. The more diverse and comprehensive this database, the more reliably the model will perform when analyzing new data. Fuzzy logic will handle the borderline cases - situations where a standard algorithm might give a definitive but incorrect answer. Instead of a rigid "defect" or "no defect" classification, the system will provide a weighted assessment that accounts for various factors of uncertainty.
A key objective is to teach the system not just to find a crack, but to evaluate its actual danger. Two defects of identical size can pose completely different risks depending on their location, the material, and the operational loads the part will endure. Integrating fuzzy logic is precisely what allows the system to consider this context, moving beyond simple threshold values.
The final outcome will be a fully integrated system ready for deployment in real-world production environments, complete with documentation, user manuals, and validated accuracy metrics. The research is supported by a grant from the Vladimir Fortov Competition for Young Scientists.
Background: The Vladimir Fortov Young Scientists Competition is held under the federal Priority 2030 program. Moscow Polytechnic University participates in this program through its strategic projects focused on a Modular Platform for Compact Electric Vehicles and the Localization of Automotive Components.
