Modulation Signal Classification Based On Centralized And Distributed For Heterogeneous Wireless Network

Modulation signal classification refers to identifying and classifying the modulation modes of all received signals to ensure accurate demodulation of signals and judge the type of transmitted signals.This technology has been widely used in military and civil fields.In recent years,deep learning has been applied to the field of communication with its excellent data expression ability,which realizes the automatic recognition and classification of modulation signals.The existing modulation signal classification based on deep learning is mostly based on centralized and decentralized,but these methods are not suitable for heterogeneous wireless network(HWN)with coexistence of subnetworks and data mismatch.Based on this,this paper carries out modulation signal classification based on centralized learning and distributed learning in HWN.The paper includes three parts:(1)Implement classification of modulation signal based on centralized learning for HWN.Convolutional Neural Networks(CNN)and Long Short-term Memory Networks(LSTM)have proven to be the main artificial neural networks for recognition and prediction problems.The paper constructs a 5-layer CNN model structure through the research of neural network,the design of objective function,the design of optimizer and the analysis of data fitting.Compared with the LSTM with the same network parameters,the 13 kinds of modulation signals in the HWN are classified by the method of central learning.Simulation results show that CNN is more suitable for modulation signal classification in HWN than LSTM,both in terms of performance and complexity.(2)Study modulation signal classification based on distributed learning for HWN.Combining deep learning-based modulation signal classification with distributed learning,a CNN with strong feature extraction capability is trained using distributed training.In distributed learning,the data of the subnetwork is kept locally,and only the model weights need to be transmitted between the subnetwork and the parameter server.The parameter server combines the model weights uploaded by the subnetwork into the global model through the model aggregation algorithm.Simulation experiments show that the modulation signal classification based on distributed learning can effectively solve the problem of data mismatch.Compared with the method based on centralized learning,it has a similar convergence speed and can greatly reduce the computational time at the expense of a small amount of accuracy.Because the size of the model weights is usually much smaller than the size of the dataset,it means that the communication overhead is much lower,allowing the modulation signal classification technique to be applied to less computationally powerful subnetworks.(3)Analyze the factors that affect distributed learning performance.The paper also discusses the convergence,acceleration and generalization of distributed learning,and studies the effects of the number of subnetworks,the number of communication rounds and the participation of subnetworks on the performance of distributed learning.The simulation results show that the performance of distributed learning will increase with the increase of the number of subnetworks.The appropriate number of communication rounds can not only save the communication cost,but also effectively improve the classification performance.In addition,when the subnetwork with poor performance and slow efficiency does not participate in the training of the global model,it can improve the performance and efficiency to a certain extent.