Registration Number
0226U004408,
(0122U002161)
, R & D reports
Title
Methods and software for interpreting machine learning models of non-parametric dynamic objects
popup.stage_title
Створення програмних засобів інтерпретації моделей машинного навчання у вигляді нейронних мереж у форму непараметричних динамічних моделей.
Head
Polozhaienko Serhii A., Доктор технічних наук
Registration Date
28-08-2026
Organization
Odesa Polytechnic National University
popup.description1
The aim of the work is to develop mathematical models of computer research of dynamic objects based on the theory of nonparametric identification, as well as methods and software tools for implementing these models in machine learning with the use of fuzzy neural networks.
popup.description2
The research aims to improve the accuracy of surrogate models by developing a method for interpreting machine learning models of nonlinear dynamic systems using nonparametric dynamic models based on Volterra integro-power polynomials, and by implementing the proposed method as software and algorithmic tools for identifying continuous systems within intelligent systems. It is shown that the use of existing surrogate models is focused on forming a qualitative picture of interpretation and does not provide sufficient accuracy in identifying machine learning models of nonlinear dynamic objects.
The relationship between nonlinear dynamic models in the form of neural networks with time delays and Volterra polynomials is studied in an analytical form. A method for estimating Volterra kernels using neural networks with time delays with activation in the form of a linear rectifier is developed. The advantage of the developed method compared to existing ones is the possibility of estimating Volterra kernels without using input signals of a special form, when the object is operating in the operating mode.
In the task of interpreting the TDNN model of a test nonlinear dynamic object, second-order surrogate models were obtained by estimating Volterra kernels according to the parameters of the trained neural network. The constructed models have an accuracy higher by 10-12% than the linear model.
A method for interpreting neural networks with time delays was developed by constructing surrogate models in the form of non-parametric dynamic models based on Volterra integro-power polynomials to describe nonlinear dynamic objects while ensuring acceptable modeling accuracy. The advantage of the developed method compared to existing ones is an increase in the accuracy of identifying dynamic objects due to the most complete consideration of nonlinear characteristics in Volterra integro-power polynomial models.
Product Description
popup.authors
Serhii M. Pelykh
Vadim F. Lozhechnikov
Volodymyr Y. Bekirov
Anatolii G. Kisiel
Liudmyla I. Protasova
Oleksandr O. Fomin
Orlov Andrii А.
Krykun Valentyn А.
Svitlana M. Hryhorenko
Viktor O. Speranskyy
Ivan Bobok
Liudmyla L. Prokofieva
Oleksandr V. Yevseiev
Dmytro V. Filin
Oleksii V. Tataryn
popup.nrat_date
2026-08-28
