Artificial neural network in the diagnostic problem of structural defects of printed components in electronic means

Problem statement. Modern radio-electronic means (RES) are complex technical systems that have found application in almost all industries and spheres of human activity. The wide functionality of RES often leads to a complication of their constructive implementation, and, as a result, to the complexity and ambiguity of diagnostic procedures performed during production and operation. In this regard, the issue of improving existing methods of technical control and developing new approaches to the diagnosis of RES in order to identify their hidden defects and increase the reliability of research results is quite acute. Goal. Improving the efficiency of diagnosing printed circuit assemblies of electronic devices in the process of their production, final inspection, testing and intended use. Research methods. At the initial stage of the study, a computer model of the printing unit under study was developed, containing detailed information about the device design. Then we analyzed the most common types of defects in printed components that occur during the production and operation of electronic devices. Seven characteristic defects were identified. Since each defect changed the type of dynamic response characteristics of the object under study, the amplitude-time characteristics of the printing unit were formed for the correct state of the device and for States with defects. Using the Monte Carlo method, a series of samples with acceptable ranges of parameter values was created for each defect. From the obtained samples (sets of amplitude characteristics of the investigated node), a fault database was formed, which was used as a comparison with the sample in diagnostic procedures. Next, a 3-layer artificial neural network (ins) was created, which was trained and tested on samples from the fault database. The results of training the ins based on activation functions allowed us to conclude that it has achieved the required level of pattern recognition and the specified reliability of the results obtained. Results. In the course of the study, a database of characteristic electronic failures was developed, for which, along with a physical experiment, mathematical modeling methods and the Monte Carlo statistical test method were used. In addition, an artificial neural network was created, which became the main tool for diagnostic research in order to detect defects in the electronic node and significantly increased the reliability of the results in comparison with existing diagnostic methods. Practical significance. To test the developed method, a series of computational experiments was performed. The type of test impact in the form of a sawtooth pulse with a linearly increasing leading edge was justified, and the parameters of this pulse were selected by calculation. The artificial neural network training technology allowed us to obtain reliable diagnostic results with a probability of P=0.99. The computational experiment was confirmed by physical tests of the radio-electronic unit on a vibration shock installation.