Optimization of Neural Network Architecture with Regard to Data Augmentation

The paper is devoted to resolving the contradiction between increasing the model's resistance to interference and distortion and complicating the task of model training under conditions of limited computational resources. The aim of the work is to determine the architecture of nonlinear dynamics models under conditions of limited training data while ensuring a given modeling accuracy. This goal is achieved by developing a method for selecting the architecture of NAS neural networks. The scientific novelty of the work lies in the further development of the method of selecting the architecture of the NAS neural network for identifying nonlinear dynamic objects, taking into account the distortions of the training dataset by adding segmented data. In contrast to the traditional approach to pre-training, the developed method allows us to build more robust models while ensuring the required accuracy. The practical significance of the work is to develop an approach to adapting the architecture depending on the augmentation methods used by developing an algorithm for selecting the architecture of a NAS neural network taking into account data augmentation, which allows building more reliable models without losing modeling accuracy. The results of experiments on modeling test objects with nonlinear dynamic characteristics are presented, and the influence of data augmentation on the quality and stability of the obtained models is analyzed. The value of the study is to determine the area of effective use of the proposed method, as tasks with a lack of labeled data in the absence of strict requirements for the speed of the modeling process