| The wireless communication signal recognition means that,in a complex electromagnetic environment,the receiver does not know which kind of modulation schemes sent by transmitter,and the receiver could recognize the modulation of the received signals.Whether in the field of electronic surveillance and electronic countermeasures,or in the civilian field we need to identify the types of the received signals,it is necessary to recognize the modulation of the wireless signal.However,with the increasing maturity and complexity of the technologies in the field of wireless communications,it is difficult to improve the performance of signal detectors merely through traditional communication theories and mathematical statistics methods.In the meanwhile,deep learning shines in computer vision and natural language processing.More and more researchers tend to apply deep learning to various fields.Therefore,some scholars realize that combining traditional communication theory with deep learning technology would have great prospects.Based on deep learning technology,this paper studies the neural network architecture and modulation recognition algorithm for wireless signals.This paper compares commonly used convolutional neural network architectures,including deep residual convolutional networks,improved convolutional neural networks,Inception,and so on.We use the Xception neural network module proposed by Google Research Institute to construct a separable convolutional neural network(SCNN)based on a deep separable convolutional neural network architecture,and use this architecture for modulation and recognition of wireless communication signals.We use the SCNN architecture and a deep separable convolution module to propose a wireless signal modulation recognition algorithm.In order to verify the accuracy of the model,this paper uses the published wireless signal dataset Radio ML2016.10 b as a benchmark for experimental testing.The experimental results show that the proposed algorithm has higher accuracy than traditional methods and general convolutional neural networks in low signal-to-noise ratio scenarios.The accuracy can reach 87.4% when the SNR is 0d B;the SNR is 2d B The above high signal-to-noise ratio scene has an accuracy rate of more than 90%,and effective modulation recognition can be performed in multiple scenes,and the recognition accuracy rate is higher than that of the traditional modulation recognizer.In addition,in response to the shortcomings of the SCNN architecture,this paper proposes a distributed architecture to implement convolutional neural network-based modulation identifiers.The distributed architecture is based on Flasks web service framework.Based on Map Reduce distributed algorithm,this paper designed the Web service of the wireless modulation recognizer with the parallel calculation and fault-tolerance.Finally,we deployed the online wireless modulation recognizer on the Alibaba Cloud server platform.The experimental results show that after the distributed algorithm is used in distributed environment.The speed of the modulation identifier on three machines is 1.5 times the response speed of the single machine,and the response speed on the five machines is 2.4 times the single machine response speed,which effectively improves the average effect speed of the SCNN architecture. |
