Mixed Data Augmentation Through the Integration of Mode Decomposition and Time-Series Transformation for Radio Signal Classification
Bin Wang, Zhuang Yuan, Jun Yong Lu, Xianchao Zhang · IEEE Transactions on Cognitive Communications and Networking · 2025
The rapid development of deep learning has provided new solutions for radio signal recognition in complex electromagnetic environments, demonstrating excellent performance and being widely applied. The data-driven based radio signal recognition methods rely on large amounts of labeled data for training. However, in non-cooperative communication scenarios, the limited availability of labeled samples can lead to model overfitting, which in turn degrades the recognition performance of the model. Therefore, this paper first combines the variational mode decomposition with time series transformation to propose a mixed data augmentation method for improving the feature extraction efficiency of the model. Secondly, considering the implementation of lightweight network structures on devices with limited computing power and storage while ensuring model recognition performance, a sparse residual network (SRNet) is designed based on a hybrid attention mechanism and sparse coding structure. Finally, comparative experiments and performance analysis are conducted on simulated and real-measured data. The simulation and experiment results indicate that the proposed method can significantly improve recognition accuracy while achieving model lightweighting.