DEEP CNN AND RESNET ARCHITECTURES FOR AUTOMATIC MODULATION CLASSIFICATION

Abstract

The paper addresses the problem of automatic modulation classification (AMC) in modern radio monitoring and cognitive radio systems. A systematic analysis of deep convolutional and residual neural network architectures for modulation recognition based on fixed-length complex IQ samples is presented. The feasibility of using one-dimensional convolutional layers is substantiated, as they enable the automatic extraction of time–phase signal patterns without preliminary construction of statistical or spectral features. Several deep neural network architectures are investigated, including a baseline convolutional model, a deeper CNN architecture, and an adapted ResNet structure with a comparable number of parameters, which ensures a fair comparison of their performance.

The simulation results demonstrate that the use of residual connections improves the stability of training deep models and provides higher classification accuracy, especially under low signal-to-noise ratio (SNR) conditions. The analysis of the confusion matrix reveals a structured pattern of inter-class errors, indicating the presence of difficult samples in the training dataset. Based on this observation, an adaptive training sample selection algorithm is proposed, which redistributes sampling probabilities according to the current classification errors. The proposed method effectively implements an importance sampling mechanism in the stochastic optimization process and focuses training on the most informative examples. Experimental results show that the use of the adaptive mini-batch formation mechanism reduces the number of training iterations by 18–25% without degrading the final classification accuracy. The obtained results confirm the effectiveness of combining deep neural network architectures with adaptive training optimization strategies for automatic modulation classification tasks under uncertain radio channel conditions.