| Automatic Modulation Classification(AMC)refers to the technology that the receiver can determine the signal type according to the information of the signal itself when the modulation type of the received signal is unknown,and it is also the basis of the non-cooperative communication system.With the increase of modulation types and the deterioration of the communication environment,it is difficult for traditional AMC algorithms to achieve effective classification,so finding an efficient AMC method has become a key problem that needs to be solved urgently.Thanks to the vigorous development of computing resources and big data,the potential of deep learning technology in computer vision,communication and other fields has been tapped.Based on this,the paper studies the combination of deep learning and AMC,and uses deep neural network to realize the integration of feature extraction and classification,so as to further improve the classification performance.The thesis contains three parts:(1)Aiming at the phenomenon that the buildings in the city are getting denser and denser,resulting in a large number of scattered signals in the wireless propagation environment,the paper simulates the environment through the Rayleigh fading model,and proposes an AMC based on the fusion of convolutional neural network and recurrent neural network.The algorithm combines the feature extraction capabilities of the convolutional neural network and the recurrent neural network with different emphasis.The recurrent neural network in the paper uses a variant that improves training speed called Simple Recurrent Units(SRU).The experiments considered three fusion structures,namely convolutional neural network series cyclic neural network(CNN Series SRU:CSS),cyclic neural network series convolutional neural network(SRU Series CNN: SSC)and convolutional neural network parallel cyclic neural network(CNN Parallel SRU: CPS),and compared with the traditional AMC algorithm.The experimental results show that the performance of the fusion algorithm far exceeds the traditional algorithm.Among the three fusion methods,CSS has a higher probability of correct classification,and the average classification accuracy under various signal-to-noise ratios exceeds 90%,which is better than the other two fusion methods.method.(2)Aiming at the common phenomenon that the moving state of the receiver in wireless communication will affect the signal frequency,the paper simulates this scenario by changing the Doppler frequency shift value of the Rayleigh fading channel,and explores the proposed method based on convolutional neural networks.Robust performance of AMC algorithm fused with network and recurrent neural network.The experimental algorithms are also three fusion algorithms and traditional algorithms.The experimental results show that with the increase of the Doppler frequency shift,the performance of the four algorithms is declining,but the performance of the fusion algorithm in this scenario is still far away.much higher than traditional algorithms.At the same time,compared with CPS and SSC,CSS has a more stable performance and can achieve more than 97% classification performance under large frequency shift and high signal-to-noise ratio,and has better robustness.(3)In view of the problem that the CSS algorithm model with excellent performance is large in size and puts pressure on storage resources,it is compressed and pruned to optimize the practicability of the model.The same classification performance and robust performance research experiments are conducted on CSS models through weight connection-based pruning.The experimental results show that when the model is pruned,the size of the model is greatly reduced,but the performance is seriously deteriorated,and the model after proper pruning can ensure the classification performance and have a smaller model size.Especially when the model pruning parameters are set reasonably,the average classification performance can be improved by nearly half a percentage point. |
PDF Full Text Request