EraseIMF: Enhancing Deep Learning-Based Few-Shot Modulation Recognition With VMD Random Erasure Augmentation

During signal transmission, errors inevitably occur, leading to bits being set to zero and affecting the signal’s integrity. To simulate this content loss phenomenon, we adapt the random erasure method from image processing and introduce it into the field of signal processing. By combining it with the time-frequency transformation method Variational Mode Decomposition (VMD), we can generate new samples to expand the training set. This method involves randomly erasing the Intrinsic Mode Functions (IMFs) components decomposed by VMD and then reconstructing the random erasured IMF components to generate new samples, which we call the EraseIMF method. By decomposing the signal, randomly random erasureing certain components, and reconstructing the signal, this method generates diverse augmented data to improve the model’s generalization ability and performance. Experiments have demonstrated that our proposed EraseIMF augmentation method performs well across different random erasure rates and various convolutional networks in few-shot scenario.