Classification and generation of soliton waves via convolutional neural networks

Abstract

This study investigates the generation, identification, and classification of soliton waves by employing deep learning methods, specifically focusing on eliminating the uncertainty and inefficiency inherent in traditional trial-and-error approaches. Initially, an extensive dataset comprising sine, square, triangular, and soliton waveforms is created using a specialized experimental setup including a nonlinear transmission line (NLTL), a signal generator, and an oscilloscope. To enhance the robustness and generalization of the deep learning models, data augmentation techniques such as flipping, rotating, scaling, and cropping are applied. Among 20 evaluated pre-trained convolutional neural network architectures, DenseNet169 exhibited the highest accuracy and is selected for comprehensive training, validation, and testing. Results demonstrated the efficacy of DenseNet169, achieving a training accuracy of 0.988, validation accuracy of 1.000, and test accuracy of 0.984. This high level of performance underscores the potential of deep learning approaches to automate and optimize soliton wave identification and generation processes reliably.