Přístupnostní navigace
E-application
Search Search Close
Doctoral Thesis
Author of thesis: Ing. Vladimír Bílek, Ph.D.
Acad. year: 2025/2026
Supervisor: doc. Ing. Jan Bárta, Ph.D.
Reviewer: The opponent will be displayed after his opinion is published.
This doctoral thesis presents a machine learning methodology for fast, accurate modeling and multi-objective optimization of electrical machines. Conventional approaches rely either on analytical equations with limited fidelity or on finite-element analyses that are computationally expensive, especially when combined with evolutionary optimization algorithms that require thousands of evaluations. To bridge this gap, Gaussian Process Regression is adopted as a surrogate model and paired with Bayesian Optimization to minimize the number of costly simulations while preserving prediction accuracy. The methodology introduces several modifications and improvements. First, it automatically selects mixed-input Gaussian Process Regression kernels suitable for the continuous and categorical variables typical of machine designs. Second, a cross-validation procedure filters outliers to curb overfitting, while an adaptive learning-curve analysis identifies the amount of training data needed for reliable predictions. Third, a MaxMin sampling routine enforces manufacturability constraints when generating candidates. Finally, three Bayesian Optimization-based workflows are equipped with custom stopping criteria and feasibility checks to guide efficient convergence. The modeling strategy is validated on two practical examples: predicting additional no-load losses in power transformers and pulse-width modulation induced harmonic losses in induction machines. In both cases, the surrogate models surpass established analytical formulas and substantially reduce computation time. Furthermore, the accuracy of Gaussian process regression models, with different kernels, was compared with other machine learning techniques and was shown to still exhibit better accuracy. The optimization workflows are demonstrated on a three-phase squirrel-cage induction motor and a spoke-type permanent-magnet synchronous motor. For the induction motor, all three Bayesian Optimization variants achieve the targeted size reduction and cost savings with far fewer simulations than genetic algorithms. For the synchronous motor, Bayesian Optimization performs similarly to a differential-evolution optimization algorithm while requiring considerably fewer evaluations. Overall, the proposed machine learning methods reduces the total number of design evaluations and shorten optimization time without compromising design quality too much. The presented methodology is readily extendable to other electrical machines or multi-physics design problems and could contribute to the development of knowledge and new directions in the design of electrical machines.
Bayesian optimization; black-box function; electrical machines; finite element method; Gaussian process regression; genetic algorithms; induction machine; machine learning; optimization; supervised machine learning; surrogate modeling; synchronous machine
Date of defence
18.12.2025
Result of the defence
Defended (thesis was successfully defended)
Process of defence
Obhajoba proběhla prezenční formou. V rámci obhajoby doktorand seznámil komisi s výsledky své disertační práce, včetně vlastních přínosů. Po skončení prezentace doktoranda jeho školitel seznámil přítomné se svým hodnocením celého průběhu studia. Následovala prezentace oponentních posudků a diskuse k dotazům a připomínkám oponentů. Poté ve veřejné diskusi vystoupili s dotazy/připomínkami k disertační práci prof. Toman, prof. Aubrecht, Ing. Hemzal, doc. Bernat, doc. Hruška., dr. Knebl. Písemný záznam dotazů je přílohou protokolu. Všechny dotazy oponentů i v rámci veřejné diskuse byly doktorandem správně vypořádány. Doktorand prokázal tvůrčí schopnosti v dané oblasti výzkumu a práce splňuje požadavky standardně kladené na dizertační práce v daném oboru. V neveřejné diskusi a po tajném hlasování komise konstatovala, že doktorand splnil podmínky par. 47 odst. 4 Zákona o vysokých školách č. 111/98 a lze jí tedy udělit titul doktor - Ph.D.
Language of thesis
English
Faculty
Fakulta elektrotechniky a komunikačních technologií
Department
Department of Power Electrical and Electronic Engineering
Study programme
Power Systems and Power Electronics (DPC-SEE)
Composition of Committee
prof. RNDr. Vladimír Aubrecht, CSc. (předseda) doc. Ing. Ondřej Vítek, Ph.D. (člen) prof. Ing. Petr Toman, Ph.D. (člen) Ing. Ladislav Knebl, Ph.D. (člen) doc. Ing. David Pánek, Ph.D. (člen) Ing. Petr Chmelíček, Ph.D. (člen)
Supervisor’s reportdoc. Ing. Jan Bárta, Ph.D.
Responsibility: Mgr. et Mgr. Hana Odstrčilová