A lightweight deep learning-based approach for automatic modulation classification using multi modal data

Abstract

Automatic Modulation Classification (AMC) is a vital task in modern communication systems, enabling efficient spectrum utilization and reliable signal demodulation. This thesis proposes a lightweight deep learning-based approach for AMC using multi-modal signal data. The methodology involves generating radio signals with eight distinct modulation types (six digital and two analog) using MATLAB’s Communications Toolbox and capturing them with ADALM-PLUTO software-defined radios. To enhance signal diversity and realism, channel impairments such as AWGN, Rician fading, and clock offsets are applied. Each signal is transformed into time-frequency representations using the Continuous Wavelet Transform (CWT), separating amplitude and phase components into grayscale scalograms. These dualmodal images are then classified using a custom dual-stream Convolutional Neural Network (CNN), designed to be computationally efficient while maintaining high accuracy. Experimental results demonstrate that the proposed model achieves 98% classification accuracy with significantly fewer parameters compared to baseline models like DenseNet, EfficientNet, and SqueezeNet. This balance of accuracy and efficiency makes the model well-suited for deployment on resource-constrained edge devices. The findings affirm the effectiveness of wavelet-based image representation and lightweight CNN design in advancing AMC performance.