| Communication signal modulation plays an important role in many fields,such as signal detection,spectrum supervision,military communication,electronic countermeasures,etc.Communication signals need to be demodulated after modulation to obtain the transmitted information,but in reality,the type of signal modulation is often not clear.Therefore,it is necessary to identify and demodulate the modulation type to obtain relevant information.Traditional communication signal recognition methods require the design of corresponding signal feature extractors to extract signal feature information,and then perform matching recognition.However,the design of this method is complex and the recognition effect is poor.Therefore,this article introduces the convolutional neural network method in deep learning to automatically extract signal image feature information,with a relatively simple structure and better recognition performance.Nine types of communication signals were generated through Matlab simulation,namely 2FSK,4FSK,2ASK,4ASK,2PSK,4PSK,LFM,8QAM,and 16 QAM.The time-frequency maps of nine types of communication signals under various signal-to-noise ratios were obtained through short-time Fourier transform,generating a total of 27000 images.1.A Dense-SE module has been designed,which achieves feature reuse through dense connections,enhances the transferability of image information,reduces the amount of network parameters and computation,and can extract important image channels in a targeted manner to improve network performance.Based on this module,the paper proposes a Dense SENet network structure,which is constructed on the Dense SE module,using Re LU and Sigmaid activation function,cross entropy loss function,and adding impulse to the random gradient descent optimization algorithm.The experimental comparison of different models under the same experimental conditions shows that this network has higher recognition performance than other methods.Compared with other four classical convolutional neural networks,it has higher recognition rate under various SNRs.The recognition rate of nine communication signals reaches 90.76% under-6d B,and the number of parameters and calculations is relatively small.It has better nonlinear expression ability and good generalization.2.A Mobile-SE module has been designed,which combines deep separable convolution,inverse residual structure,and channel attention mechanism.While significantly reducing parameter and computational complexity,it can also focus on extracting prominent channel features,enhancing the module’s feature extraction ability.Based on this module,a Mobile-SENet network structure is proposed in the paper.After the verification of the experimental comparison of different models under the same experimental conditions,this network has higher recognition performance compared to other methods and has good recognition rates under various signal-to-noise ratios.The recognition rate still reaches 85.36% at-6d B,which is higher than other classic networks.It is not significantly different from deep network ResNet,and its parameter quantity is only 1/6 of Res Net,and the computational cost is 1/12 of Res Net,indicating good real-time performance.3.A Fused Mscale module has been designed,which reduces the computational complexity of the model and improves its efficiency and accuracy performance by integrating deep separable convolutions,channel attention,and multi-scale mechanisms.On the basis of this module,the Efficient Scale Net network was introduced.After the experimental comparison of different models under the same experimental conditions and comparison analysis,the network showed high recognition performance under different signal-to-noise ratios and signals,and also had a higher recognition rate compared to other methods.The recognition rate reached 92.52% at-6d B and 98.62%at 9d B,with relatively low computational complexity.It can provide more degrees of freedom to a certain extent,making the model more flexible and efficient. |
