Deep Unfolding Networks for Robust Modulation Classification in Cognitive Radio Systems

Abstract

The modulation classification is essential in cognitive radio (CR) systems as a way of providing dynamic spectrum access as well as the enhancement of the spectral efficiency. Nevertheless, current machine learning and deep learning methods are usually susceptible to noisy and rapidly varying setups especially at low signal-to-noise ratio (SNR levels). The given paper proposes a new architecture of Deep Unfolding Network (DUN) for Automatic Modulation Classification (AMC) in CR systems. The given technique is able to unfold a typical iterative inference algorithm into a structured training neural network and therefore inherits both interpretability of model-based signal processing and versatility alongside learning capabilities of data-driven models. The performance of DUN is tested on RadioML2016.10a dataset on different modulations coupled with SNR and in realistic wireless environments by simulating fading, multipath and IQ imbalance. Findings illustrate that DUN is substantially more accurate, robust and computationally efficient than conventional CNN and LSTM-based classifiers under low-SNR conditions (below 0 dB). Also, the network can converge faster with fewer parameters, making it easy to deploy, even in resource-limited CR hardware. This paper makes a step towards defining DUN as an efficient and scalable real-time AMC tool in intelligent wireless systems and towards more adaptive and resilient cognitive radio frameworks.