Baida V. Methods and information technology for diagnosing failures in semiconductor power converters based on digital twins

The dissertation solves an important scientific problem of improving the effectiveness of online fault detection and identification in semiconductor power converters by developing new methods and an information technology for diagnostics based on digital twins, which provide automated online diagnosis of component defects without the need to measure the operability of individual elements of the device. The proposed approach combines metaheuristic optimization for identifying the parasitic parameters of components as constituents of the mathematical model (calibration of the digital twin to a specific device unit), qualitative analysis of output signal deviations for fault localization, and remote laboratory technologies for experimental verification. The first chapter analyzes current approaches to the diagnostics of electronic devices, in particular methods based on built-in self-test, signature analysis, and machine learning. The concept of digital twins as a diagnostic tool is considered, and methods for constructing digital twins and identifying the model parameters of power electronic converters are analyzed. The current state of remote laboratory technologies and their potential for verifying diagnostic systems are investigated. The main limitations of existing approaches are identified. The research goal is formulated and the research tasks are stated. The second chapter develops a new integrated method for online diagnostics of semiconductor power converters based on a digital twin. The method is based on representing the input and output parameters of the device as vectors and forming a correspondence table between the deviations of the output signals from their reference values and the possible causes of faults. The reference values are obtained using a digital twin calibrated to the specific device unit. An approach with discrete deviation states is proposed, which provides robustness to measurement noise and low computational resource requirements during operation, since all resource-intensive computations are performed at the system initialization stage. Four main methods for generating correspondence tables are systematized: systematic simulation, analytical derivation, formalization of expert knowledge, and the experimental method, as well as a hybrid approach, which, based on the results of a qualitative comparison, is recommended for practical applications. The third chapter substantiates the choice of a genetic algorithm for solving the problem of identifying the parasitic parameters of a DC-DC power converter in order to calibrate its digital twin. It is proposed to determine the reference signal for calibration by the criterion of equidistance of its indicators from the boundary values of the experimental sample, owing to which the deviations of the calibrated model are symmetric and are used directly as the initial diagnostic thresholds. A comparative analysis of gradient-based methods, particle swarm optimization, the grey wolf optimizer, and the genetic algorithm is carried out. The mathematical model of the identification problem is formulated as a problem of global minimization of the deviation between the outputs of the model and the real device. An approach with logarithmic gene encoding is introduced, which ensures efficient handling of parameters spanning up to 9 orders of magnitude. The method is validated on a DC-DC converter with the identification of seven parasitic parameters based on experimental measurements. The fourth chapter further develops technological approaches to designing the architecture of remote engineering laboratories: a method for multi-criteria comparative analysis of the information technologies used to build them is devised, which makes it possible to move from an empirical to a scientifically substantiated choice of technology platform. A system of six criteria with substantiated weighting coefficients is proposed: technology level, updates and support, LMS integration capability, architectural simplicity, real-time performance, and ease of configuration. The fifth chapter presents the experimental validation of the proposed methods. A method for comparing backend information technologies for interaction with LabDiscoveryEngine is developed, which provides a quantitative justification for the choice of the software platform for implementing the remote laboratory. A hardware-software complex of a remote power electronics laboratory based on the LabDiscoveryEngine technology for semiconductor DC-DC converters is developed. The obtained results are of practical significance, since the proposed information technologies and methods can be used to build systems for automated diagnostics and predictive maintenance of power electronic converters, as well as a baseline architecture for creating remote laboratories in the field of electronics.