A Consistency Regularization Training Method for Automatic Modulation Classification under Incomplete Information

Deep neural networks provide innovative solutions for automatic modulation classification. However, the performance of deep neural networks (DNNs) is significantly affected by the incomplete signal pattern conditions, which are caused by the limited number of training samples and the diversity of modulation signals. Existing deep learning models typically utilize regularization methods to enhance the model’s generalization and robustness. However, these methods suffer from the issue of inconsistent training and testing output distributions, which can lead to a decline in model performance during testing. To address this issue, a consistency regularization training method is proposed for AMC under incomplete signal pattern conditions. The method utilizes the designed consistency regularization loss function to explicitly introduce regularization constraints between the submodel outputs and the full model outputs, thereby mitigating the inconsistency in output distributions caused by Dropout and ultimately enhancing the model’s generalization capability. Experimental results demonstrate that our method effectively improves the performance of DNNs in AMC tasks under incomplete signal pattern conditions.