Automatic Modulation Classification (AMC) is an essential element of cognitive radio network that provides an awareness of the spectrum as well as adaptive communication. Nevertheless, the current deep learning-based AMC schemes frequently struggle to achieve a trade-off between classification precision, computational speed, and model explainability, which are essential to real-time execution on edge computing devices. To solve these problems, this paper presents a wavelet-augmented, pruned, and quantized convolutional neural network (CNN) framework that uses Grad-CAM explainability. Evaluations on the RadioML2016.10a dataset across SNR levels from −20 dB to +20 dB are done. The suggested model obtains 99.1% accuracy at +10 dB. Although maintaining 85.6% accuracy at −4 dB SNR, SNR shows robustness under noisy surroundings. With an average inference latency of 5.2 ms and a power consumption of 5.2 W, deployment profiling on an NVIDIA Jetson Nano with a USRP B210 confirms suitability for real-time cognitive radio applications. Transparency and trust are increased when Grad-CAM visualizations verify that the model takes into account modulation-specific characteristics like envelope variations in AM signals, amplitude clusters in 16-QAM, and phase transitions in QPSK. Compared to more recent AMC approaches (2016-2024), the proposed framework is identified as the first to be both highly accurate, lightweight edge deployable, and explainable, providing a solid base to practical AMC of next-generation wireless systems.
Ahmed et al. (Tue,) studied this question.