A Hybrid Approach for Cross-Dataset Modulation Recognition of Wireless Interference

The cross-dataset problem in modulation recognition that may arise from some practical factors, such as unknown channel environment and the transmitter radio frequency characteristics, can severely reduce the recognition accuracy. To tackle this challenge, we propose a hybrid approach that leverages the cross-attention mechanism to combine the received signal’s IQ features, the statistical features, and the transform-domain features to improve recognition accuracy. Specifically, we first introduce multiple delay vectors in the cyclic cumulant (CC) and exploit their varying sensitivities to different channel environment and transmitter characteristics to improve robustness in cross-dataset scenarios. Furthermore, the proposed hybrid approach fuses in-phase and quadrature (IQ) features and time-frequency (TF) graph features with the improved CC features, where IQ features and TF graph features enhance the recognition accuracy, while CC features ensure the robustness in cross-dataset scenarios. In addition, the proposed hybrid approach introduces a flexible design that utilizes transfer learning pretraining on the large dataset which applies few-shot learning on the new dataset to adapt to varying data lengths while ensuring the recognition accuracy. Simulation results verify that the proposed solution achieves better recognition accuracy than baselines.