Виявлення дефектів у періодичних структурах

• Дата: 2025
• Автор: Zhydkov, Volodymyr
• Формат: Стаття
• Мова: English
• Опубліковано: V.M. Glushkov Institute of Cybernetics of NAS of Ukraine 2025
• Теми: регулярні структури; дефектоскопія; обробка зображень
• Онлайн доступ: https://jais.net.ua/index.php/files/article/view/437
• Назва журналу: Problems of Control and Informatics

Rejection of defective parts and elements is an important part in quality improvement, maintenance and accident prevention. Some imaging instrumental methods of test and inspection have improved ability to reveal defects, but currently have disadvantages, such as low accuracy of defect classification and need of human operator. Therefore, there is a high demand for a solution that is reliable enough, can exclude human element and mathematically grounded unlike probabilistic and neural network models. The paper presents generally applicable algorithmic solution for a system that can analyze images and reveals defects automatically on par or better than a human operator. The general principle is based on approach that some general properties about the analyzed part (periodicity) is known beforehand. The method of defects reveal by pattern-recognizing not defects themselves but periodic background is proposed. Mathematical formulation for an error function adapted for approximating images and data with defects is proposed, and viability of the functions analyzed, including computational experiment to verify its validity. Using ab initio considerations, a new type of error function is suggested, it specially tailored to approximating datasets with high density of outliers, with performance comparable and in certain situations exceeding -norm in that respect. Thus, the algorithm suggested, by using which image of the object examined can be reconstructed in its «idealized», defect-less state and, further by comparing image to the reconstruction, the defects can be revealed. Furthermore, statistical model used in formulation of the error function allows estimation of defect probability in the area based on analysis of deviation and measurement errors. The approach is demonstrated and tested on a practical example. All essential stages of the algorithm, such as preliminary analysis, and model building are outlined. It is demonstrated capability of revealing defects without human operator intervention and without machine learning, only having basic information about object investigated.