Research On Communication Signal Modulation Recognition Technology Based On DAE＿Transformer

Signal modulation recognition has been one of the research hotspots in the field of deep learning.It has important research value in the fields of cognitive electronic warfare,communication countermeasure and non-cooperative communication.In some complex scenarios,the research of modulation recognition of communication signal is still challenging.In the absence of cooperation,accurate and effective identification of the modulation mode of the target signal is a prerequisite for signal demodulation and processing.Deep Learning has had remarkable success in a wide range of fields,from natural language processing and computer vision to economics and bioinformatics.Recently,DL may be studied for signal classification and modulation recognition tasks,in which automatic powerful feature learning capabilities enable it to achieve higher accuracy.In order to improve the accuracy of modulation recognition of communication signal,the following research work is done in this paper:Firstly,a novel modulation classification framework based on DAE＿Transformer is proposed.The autoencoder extracts the low-dimensional representation of the input data in an unsupervised manner,stably capturing the salient features of the high-dimensional data while performing dimensionality reduction.In order to solve the problem of large differences within modulated signal classes,this paper embeds the multi-head attention mechanism in the transformer structure in the denoising autoencoder recognition model.With multiple sets of linear projections,efficient transformations of Query,Key,and Value can be achieved without relying on a single attention set.In addition,all conversion results are sent synchronously to the attention collection to achieve goals more efficiently.With the increase of the number of layers of network structure,the objective function is also easy to fall into local extremes,resulting in the difficulty of convergence of the parameters of the model during the training process,and the accuracy of the final model will decrease.At the same time,the gradient vanishing problem will be more serious,so in the training process,the introduction of residual network,which can greatly improve the depth of the network that can be effectively trained,further improve the robustness of the model,and enable it to better deal with the gradient vanishing problem.The experimental results of this paper on the public dataset Radio ML2018.01 A show that when the signal-to-noise ratio is 20 d B,the recognition accuracy of the algorithm for the 24 modulation methods reaches 96.14%,which effectively improves the recognition accuracy of the modulated signal compared with the existing recognition methods.Finally,from the perspective of classification and identification of actual RF signals,combined with the idea of software radio,using corresponding hardware such as Lab VIEW and the universal software radio platform USRP,through building a semi physical simulation model,communication interconnection in a real environment is realized,and communication signals in a real signal environment are collected,analyzed,and automatically identified,achieving good results.