Volkov D.D.Acoustic electrodynamic transducer modelling methods – Qualifying scientific work on the rights of the manuscript. Thesis for the degree of Philosophy Doctor, in specialty 171 “Electronics”. – National Technical University of Ukraine "Igor Sikorsky Kyiv Polytechnic Institute Kyiv, 2022. The thesis is dedicated to the study of linear and nonlinear models of acoustic electrodynamic transducers with the aim of improving existing and developing new, more accurate and convenient methods using modern approaches, such as: optimization methods, genetic algorithms and artificial neural networks. The work consists of four main sections. The introduction describes the relevance of the problem and provides an overview of modern modeling methods, the second part describes in detail the linear models of converters, including the classic Thiel/Smoll model and the standard method of finding its parameters - the added mass method. New methods for more accurate and flexible finding of linear model parameters are proposed, such as the method of Bl parameter brute-force search an the genetic algorithm application. The proposed methods were compared with the classical added mass method and the advantages and disadvantages of different methods were discussed. In the third section, the nonlinear transducer models are considered. A statespace nonlinear loudspeaker model is derived and its transformation to the canonical form is demonstrated. Also, a fundamentally new approach to modeling electrodynamic transdurcers using recurrent neural networks is proposed. Both methods are compared with each other on the basis of practically measured data and conducted experiments. The last section gives general conclusions from
the work performed. As a result of the work, it was possible to obtain the following new results: 1. The latest fully automated method for finding the loudspeaker’s force factor Bl without the influence of other model parameters using exclusively measured values, such as: voltage at the terminals, current through the voice coil and displacement of the moving part of the transducer, is proposed. Such a method can be considered the most accurate indirect method of measuring Bl at the moment. 2. Based on the force factor Bl identification approachd, amethod for separating the mechanical and electrical impedances of an electrodynamic transducer is proposed using also only measured values to achieve maximum accuracy. As a result, these impedances can be considered separately from each other, which opens up more opportunities for their research and development of more accurate models. 3. For the first time, the genetic algorithm was applied to find parameters of a linear model of an electrodynamic transducer and its practical comparison with the classical method of added mass was carried out. 4. On the basis of the genetic algorithm, a universal generalized structure is proposed for finding parameters of arbitrary electrodynamic transducers models of, which greatly facilitates and accelerates the process of their research. 5. For the first time, a recurrent neural network was used to model nonlinear behavior of electrodynamic transduers. The complete process of training and testing this neural model is presented. 6. A practical comparison of the nonlinear model using a recurrent neural network with the most widely used model in the industry - nonlinear state-space model - is carried out. The practical significance of the obtained results lies in increasing the accuracy and facilitating the identification of linear and non-linear models of electrodynamic transducers. The presented methods can be used directly in the industry as well as for research purposes for a deeper analysis of the loudspeakers behavior, namely: the possibility of rapid adaptation of the genetic algorithm to more complex models with a larger number of identified parameters without loss of performance and rapid convergence of the algorithmwas demonstrated. This shows the universality of the genetic algorithm and the possibility of its use for more accurate models that are difficult to identify by classical methods. For modeling the nonlinear behavior of electrodynamic transducers, the appliction of recurrent neural networks was proposed for the first time. This approach allows to quickly find a model of a "black box"type that can be directly
used as a digital double of a modeled loudspeaker as a component of more complex systems. The proposed model was compared with the most widely used in the industry nonlinear model of the electrodynamic converter in the state-space based on the analysis of the measured responses in temporal and spectral representations.
