Enhanced Automatic modulation classification Using Leveraging Semi-Supervised Learning

Automatic modulation classification (AMC) is essential for identifying the modulation scheme of received signals, a key step bridging signal reception and demodulation. This process traditionally relies on prior signal information. Deep learning’s role in AMC has been notable, though its efficacy hinges on the volume of labeled data. Addressing the scarcity of labeled samples, our research introduces a semi-supervised AMC approach using consistency regularization and pseudo-labeling. This method harnesses unlabeled data’s distributional information to mitigate the shortfall of labeled data. We designed a dual-component objective function: one segment calculates loss from labeled data training, while the other assesses regularized loss from augmented unlabeled data. These combined losses concurrently refine the model parameters. Comparative tests reveal our method’s superiority over established algorithms like decision trees, support vector machines, pi-models, and virtual adversarial training, demonstrating enhanced performance in AMC.