双骨干轻量抗噪的无人机射频信号识别模型OA

To address the problems of low recognition accuracy and high model complexity of unmanned aerial vehicle(UAV)radio frequency(RF)signals recognition in complex noise environments,a lightweight heterogeneous dual-backbone convolutional neural network(HDB-CNN)is proposed.This model takes the time-frequency graph generated by short-time Fourier transform(STFT)as input.The efficient convolution backbone adopts channel expansion and residual connections to enhance the capture ability of short-term abrupt features.The densely-connected backbone explores long-term frequency band correlations through layer-by-layer feature reuse and multi-scale pooling.The features of the two backbones are concatenated and fused to achieve complementary time-frequency features,and large-size global average pooling(GAP)is combined to compress parameters and reduce overfitting risk.Compared with MobileNet v3,the model reduces the number of parameters by 57%,enabling its suitability for edge device deployment.In experiments based on the DroneRFa dataset,HDB-CNN improves the average accuracy by 7.8%compared with ResNet50 under Gaussian white noise,and by 8.9%under-9 dB mixed Gaussian white noise and impulse noise.