Novosibirsk State Technical University has developed an intelligent quality control system for industry. The technology allows you to automatically find cracks, dents, and corrosion spots on a steel surface based on photos from a conventional camera.
The system developed by students of the Faculty of Automation and Computer Engineering is based on a triplet neural network that does not need thousands of ready-made images. The system works effectively with a small number of training examples. Several photos of each type of defect are sufficient for analysis, even if they were taken in low light and at different scales.
"We have created a tool that is able to quickly adapt to new, rarely encountered types of damage without lengthy and expensive data re-labeling. To do this, we used architecture, which learns to "understand" the essence of the defect, and not just memorize pictures," said Egor Antonyants, project manager, assistant Professor at the Department of Automated Control Systems at NSTU-NETI.
Automated visual control systems based on classical algorithms, as a rule, require ideal shooting conditions. To train most modern neural classifier networks, it is necessary to search for huge amounts of labeled data. The uniqueness of the NSTU-NETI development lies in the fact that it offers a compromise — high accuracy with minimal preparatory work with data.
The intelligent quality control system created at Novosibirsk University shows high efficiency with limited data. Based on the test data, the system demonstrated an error detection accuracy of more than 87%, which significantly exceeds the results of traditional machine learning methods based on manual description of features. According to the developers, this makes it a profitable solution in enterprises where collecting thousands of marriage examples is difficult or economically unprofitable.
"The technology is intended for implementation in quality control and predictive maintenance systems at industrial enterprises, primarily in metallurgy and mechanical engineering. It will automate the process of monitoring steel surfaces, determine the need for equipment maintenance based on early signs of wear and improve the overall reliability and safety of production lines. In the future, the system can be adapted to monitor the condition of bridges, pipelines and other structures where uninterrupted operation is critically important," said Egor Antonyants.
The system has already passed a number of tests. According to Vitaly Zaozernov, one of the main developers of the project, a third-year student at the Faculty of Automation and Computer Engineering, tests on a public database of images of steel defects showed high accuracy in recognizing various types of damage, which confirmed the practical value of the approach used.