A simple data augmentation method for automatic modulation recognition via mixing signals

Automatic Modulation Recognition(AMR) is crucial in cognitive radio, transitioning from traditional methods to deep learning-based classification. However, deep learning requires large datasets for effective training, making data augmentation vital for enhancing model generalization and accuracy. It helps to expand the dataset, thereby improving the models’ generalization ability and accuracy. We proposes a novel data augmentation approach for AMR task based on signal mixing, with four specific methods designed to improve the performance of Mixed Signals. The experimental results demonstrate that the proposed methods lead to a notable improvement in the classification accuracy of deep learning-based AMR models in diverse scenarios. Notably, experiments conducted with datasets at a single signal-to-noise ratio show a significant increase in classification accuracy, confirming the effectiveness of the signal mixing method.