Momot A. Improvement of the method for defining defects characteristics of multilayered materials by active thermal testing

The thesis describes that in the present stage of methods of thermal non-destructive testing development important task is not only to identify and determine the coordinates and transverse dimensions of defects, but also to measure their depth and thickness. The factors that influence on results of thermal testing are analyzed and the nature of relationship between informative parameters is described. Traditional mathematical and statistical methods of thermal defectometry are considered and their disadvantages are established. In addition, considered traditional methods of thermogram processing do not allow the automatic classification of defects by type and to determine their thickness. The thesis describes the possibilities of using artificial neural networks for improvement of defect characterization methods. Features of construction the neural network systems for solving tasks of defects classification and their depth and thickness determination are considered. The performance of neural networks and traditional methods of thermogram processing are compared. Advantages of neural networks over traditional algorithms are shown. It is shown that in known literature the tasks of simultaneous defects classification by type and determination of their depth and thickness are not solved; the methods of determining defects depth or their thickness by solving the regression task using neural networks or traditional methods have not been investigated; the task of constructing thermal tomograms of object of testing internal structure is not solved. Research aim in the form of development of a neural network automated system of thermal fields complex analysis, which will have higher efficiency in comparison with systems based on traditional methods of thermogram processing is formed in this thesis. In order to improve the methods of thermal defectoscopy and defectometry and automate data processing, a subsystem of digital thermogram processing consisting of three neural network modules has been substantiated and developed in the thesis. Possibility of using feedforward backpropagation multilayer neural networks with fully-connected layers in defects detection and classification module and modules of determining defects depth and thickness is described. Algorithms of formation of training datasets for tasks of defects classification and determination of their depth and thickness are formed. The procedure of neural network modules training is described and the corresponding software is developed in MATLAB. Software implementation of virtual devices, which embodies algorithms of neural network modules and results post-processing is implemented in NI LabVIEW software. A graphical user interface for an automated thermal field analysis system has been created. That includes controls, defectometry tools and blocks for graphically displaying information about defects location and the internal structure of object of testing. The computer simulation of the process of active thermal testing of a multilayer carbon fiber specimen with artificial internal defects is carried out. According to results of efficiency evaluation of obtained thermogram sequences processing by different methods, it is established that the developed neural network system provides the highest indicators of quality of defect classification and defectometry accuracy among considered methods. The influence of neural network architecture on performance of neural network modules of developed system in case of computer simulation data is investigated. Experimental researches have shown that developed system allows to carry out error-free detection and classification of defects by type. Estimation of defects depth and thickness using developed system was conducted with the maximum error of ± 3.19 % and 3.50 % respectively. It is proved that developed system has higher reliability of testing and accuracy of defectometry in comparison with traditional algorithms even in conditions of uneven heating. Based on the results of research, recommendations on methodology of testing using developed automated system are formulated.