A Research On Signal Recognition Based On Lightweight Network

Communication Signal Modulation Classification is a technology that can recognize the modulation type of the unknown communication signal received,and it is also the key basis for the next step of signal analysis.With the development of artificial intelligence technology in the past few years,the application of deep learning technology to modulation recognition has become a mainstream trend.However,the complexity and computational complexity of the deep model are large,and it is difficult to meet the requirements in a complex environment.Therefore,this paper studies the lightweight technology of the modulation signal depth model,and the main results are as follows:1.A light-weight algorithm for signal recognition model based on channel pruning is realized.This method uses the scale factor value of the batch normalization layer of the model as the evaluation standard,and deletes the redundant channels in the model based on the global channel pruning ratio of the model.The results show that the signal recognition model channel compression method proposed in this paper can compress the parameters and complexity of the model without reducing the accuracy of the model.2.A light-weight algorithm for signal recognition model of mixed pruning of chan-nels and layers is proposed.On the basis of channel pruning,this method takes the propor-tion of the remaining channels in each layer as an evaluation criterion,deletes redundant convolutional layers in the model,and further reduces the amount of network parameters.The results show that this method can simultaneously reduce the number of channels of the convolutional layer and the depth of the model within a small range of accuracy loss,and reduce the overfitting problem of the model within a certain range.3.From the perspective of parameter quantization,the parameters of the signal recog-nition model is quantized from floating point to integer.This method can convert most of the floating-point operations of the model to integer operations within a small range of accuracy loss,thereby improving the inference speed of the signal recognition model.4.From the perspective of parameter binarization,a method of training binarized signal recognition model is realized.The method completes the parameter quantization of the model during the training process,and combines the characteristics of parameter quantization to adjust the order of the network model and optimize the structure,which reduces the accuracy loss caused by parameter binarization.The results show that model binarization can better compress the size of the model and reduce complexity within a small range of accuracy loss.5.Taking the raspberry pi embedded platform as an example,the lightweight net-work modulation recognition of embedded devices is realized,and experiments are car-ried out to compare the complexity of the model before and after lightweight,and verify the lightweight performance of the algorithm.The above work has been verified by experimental simulations,which can build a signal recognition model based on a lightweight network,reduce the number of trainable parameters and algorithm complexity of the signal model,and reduce the signal model's inference time and resource consumption in equipment with limited computing resources and finally expands the application scenarios of using the depth model for modulation signal recognition.