自动调制识别是现代通信系统中一项重要技术.为提高通信系统对不同调制信号间的识别性能,文中首先探索了包含11类调制信号的公开数据集RML2016.10A上原始同相正交(In-phase and Quadrature,IQ)格式数据和经过数据预处理后的幅度和相位(Amplitude and Phase,AP)格式数据的差异;随后,依据原始IQ格式数据和AP格式数据在特征提取过程中对局部相关性及时序特征敏感性的差异,设计了针对空间特征提取的SFE-Block模块、针对长期依赖关系提取的TFE-Block模块,以及联合时空特征提取模块STFE-Block,并将前两者的输出特征作为STFE-Block模块输出特征的重要补充进行特征融合,以全连接(Fully Connected)层负责最终分类.实验结果表明,本模型在数据集RML2016.10A上表现良好.当信噪比(Signal to Noise Ratio,SNR)低于-8 dB时,平均识别精度比其他模型提升7％,而SNR在0～18 dB时,平均识别精度比其他模型提高1％～8％,且在SNR为16 dB时,最高识别精度达92.95％.此外,在RML2016.10B数据集上重复了实验以检验模型泛化性,所得结果同样最优,且当SNR为12 dB时,最高识别精度达到 93.6％.
Automatic Modulation Recognition Method Based on Efficient Dual Stream Input Structure
Automatic Modulation Recognition(AMR)is a critical technology in modern communication systems.To improve the recognition performance of communication systems for different modulation sig-nals,this study initially explores the differences between the original In-phase and Quadrature(IQ)for-mat data and Amplitude and Phase(AP)format data after preprocessing.Subsequently,based on the sensitivity of local correlation and temporal features during feature extraction from both IQ and AP format-ted data,we designed the SFE-Block module for spatial feature extraction,the TFE-Block module for ex-tracting long-term dependencies and the STFE-Block module for extracting joint spatiotemporal features.The outputs of these modules supplement the joint spatio-temporal feature extraction provided by the STFE-Block module,which are then fused for feature integration,with a fully connected(FC)layer re-sponsible for the final classification.Experimental results demonstrate that our model performs well on the public dataset RML2016.10A.When the Signal to Noise Ratio(SNR)is below-8 dB,the average rec-ognition accuracy of this model improves by 7％compared to other models.In the SNR range of 0～18 dB,the average recognition accuracy exceeds that of other models by 1％～8％,and achieving a maxi-mum recognition accuracy of 92.95％at a SNR of 16 dB.In addition,the experiment was replicated on the RML2016.10 B dataset to test the model's generalizability,where it also showed favorable results,reaching a peak recognition accuracy of 93.6％at an SNR of 12 dB.
万方数据
维普
