Performance analysis of different signal representations and optimizers for CNN based automatic modulation classification

Abstract

Automatic Modulation Classification (AMC) plays a crucial role in non-cooperative communication systems by identifying modulation types of received signals without prior information. Recently, Convolutional Neural Networks (CNN) based AMC techniques have shown great promise in achieving high classification accuracy for multiple modulation schemes. In this regard, researchers have used different input signal representations and optimizers for training CNN models. This paper investigates the effectiveness of using constellation diagrams and spectrograms as input representations along with various optimizers for CNN based AMC using SqueezeNet model. The optimizers assessed include Stochastic Gradient Descent with Momentum (SGDM), Adaptive Moment Estimation (ADAM), and Root Mean Square Propagation (RMSprop). Eleven modulation schemes were analyzed, and classification performance was measured using accuracy, F1 score, and specificity. The best results were obtained with constellation diagrams for RMSprop optimizer, achieving an accuracy of 99.55%, an F1 score of 99.54%, and a specificity of 99.95% at 20 dB SNR. The ADAM optimizer with constellation diagrams followed closely, yielding 96.81% accuracy, 96.82% F1 score, and 99.63% specificity. These configurations also provided the best results over wide range of SNRs from − 20 dB to 20 dB. Additionally, a comparative analysis with previous approach utilizing IQ sequential signal frames demonstrated a significant improvement in classification accuracy with the mentioned configurations. These findings highlight the effectiveness of specific input types and optimizers in enhancing classification performance for CNN-based AMC applications.