Modulation Recognition Based On Instantaneous Features Fusion And MLP-Mixer

With the rapid development of 5G technology,many innovative technologies and various standards have emerged in the field of wireless communication.As the number of devices available to improve system performance increases,the amount of data increases exponentially,and the existing algorithms are not efficient in processing large amounts of data,resulting in a large amount of wasted resources.To solve these problems,Cognitive Radio(CR)has become an important application in the field of wireless communication.However,these emerging applications require models with higher accuracy and lower computational time.Modulation Recognition(MR)is one of the important technologies in communication systems,the main purpose of which is to identify the modulation mode of the received signal to overcome the influence of noise interference.However,due to the negative effects of noise and multipath fading conditions,as well as the increasing number of advanced modulation types,improving the recognition accuracy at low signal-to-noise ratio has been a challenging problem.Therefore,aiming at the problem of modulation recognition accuracy under low signal-to-noise ratio,this paper proposes an algorithm that fuses instantaneous features and uses improved MCLDNN framework for classification.Firstly,the parameters and generation mode of the required modulated signal data set are studied,and the corresponding data preprocessing operation is done.Then we study the different instantaneous features we consider to extract,and calculate the instantaneous characteristic values of each type of modulation signal as the input of the network model.Then an improved network framework is used to improve the recognition accuracy under low signal-to-noise ratio.The experimental results show that compared with CNN and CLDNN network models,the proposed method improves the overall recognition accuracy to 93.3%.On the other hand,when the signal-to-noise ratio is 0dB,the fusion of two feature values has the best recognition effect compared with the fusion of one feature value and four feature values,and the accuracy can reach 88.4%,which is 2%higher than that without the fusion of feature values.Aiming at the problem of the large number of parameters in the network model,an MLP-Mixer framework was proposed in order to reduce the computing burden of the network.Firstly,the structure of the multi-layer perceptron and the MLP-Mixer were analyzed and introduced,and the data were preprocessed to be suitable as the input of the MLP-Mixer.Then,the MLP-Mixer method was compared with CNN and MCLDNN,which greatly reduced the number of parameters.Then,the model parameters of the MLp-mixer frame were changed to add different amounts of MLP blocks,and different size patches were adopted respectively.Finally,the modulated signals are classified using the trained model.The simulation results show that when the input data shape is(2,128,1),the number of parameters in the MLP-Mixer model is nearly 10 times less than that of CNN and MCLDNN,and the recognition accuracy is 10%higher than that of CNN,but less than that of MCLDNN.If the input data shape is(2,128,128),the modulation recognition accuracy is up to 88%.Moreover,when the number of MLP layers added increases,the recognition accuracy can be improved,while the increase of patch blocks will reduce the recognition accuracy.By training the model on two different data sets,the proposed model is robust.