Automatic Modulation Recognition for Radio Frequency Proximity Sensor Signals Based on Masked Autoencoders and Transfer Learning

Automatic modulation recognition (AMR) for radio frequency proximity sensor (RFPS) signals is an essential prerequisite for performing corresponding jamming strategies as countermeasures. However, realizing AMR for RFPS signals in the interrupted sampling mode involves great challenges, including truncated RFPS signal reconstruction and efficient learning with small sample sizes and low signal-to-noise ratios (SNR). Therefore, this article proposes an AMR method based on masked autoencoders (MAEs) and transfer learning. First, a denoising model based on the encoder–decoder structure of the MAE is constructed to improve the SNR of interrupted sampling RFPS signals. Second, an AMR model based on time–frequency (TF) reconstruction and transfer learning is constructed. This model autonomously learns the internal distribution law of the truncated signals through TF reconstruction process pretraining and a random mask structure, reducing its sample size dependence. Finally, the pretrained model is migrated to an actual RFPS signal dataset acquired in the interrupted sampling mode through transfer learning for fine-tuning, thus realizing accurate RFPS signal recognition in the interrupted sampling mode under a small sample size. With the method proposed in this article, the recognition accuracy of interrupted sampling RFPS signals beyond 95% when the SNR exceeds −14 dB.