Research And Application Of Key Technologies Of Signal Detection And Recognition Based On Convolutional Neural Network

To deal with the complex problems of signal detection and recognition,which are difficult to be solved by traditional means,this paper studies the combination of these problems and deep learning technology.Furthermore,this article studies the application and deployment of signal modulation recognition network.The main contributions and innovations are as follows:1.For the problem of signal detection in a real scene,this paper designs a model based on a deep neural network.Firstly,aiming at the signal detection problem in this scene,it is abstracted as a binary classification problem in deep learning.Then,the modulation recognition dataset with sophisticated channel modeling is used as the data received by the receiver,and white noise with different power is generated at the same time.The training dataset and test dataset for this method are generated by the root raised cosine filter with random roll-off coefficient using the previously generated data to simulate the actual scene.Furthermore,the basic structure of the network is proposed,and the variables involved in network structure design are abstracted into 9hyperparameters,and the quasi-optimal model under the current dataset is found using hyperparameter search technology and the dataset generated before.Finally,the generated test data are simulated in the optimal environment.The test results show that compared with the traditional energy detection method,the proposed network can more accurately distinguish the signal and noise,the accuracy raising from 72.76% to 97.09%,which provides a feasible solution to the problem of signal detection for the actual scene.2.To overcome the drawbacks of communication signal modulation recognition,a brand-new combined deep neural network layout is presented.The proposed design can acknowledge 24 kinds of modulations consisting of public datasets.The suggested network consists of a deep network for identifying 24 inflection settings as well as a shallow one for identifying two modulation modes,which minimizes the mistake triggered by the larger network.The simulation outcome on the public dataset reveals that the deep network can acknowledge signals besides AM-SSB-SC much more properly,and the major recognition errors are lessened utilizing the AM-SSB-WC as well AM-SSB-SC signals.When the signal information is identified as these two modes of modulation settings,they will certainly be re-recognized by the superficial network to get the recognition outcomes.After the shallow network is included,the overall recognition accuracy of the mixed network is discovered to be at 98.7%,higher than the original best result 96.3%.This recognition accuracy is discovered to be the most effective amongst current works published using this public dataset.3.One of the major drawbacks of the GPU deployment using separate convolutions in the proposed networks lies in their complexity and flexibility.To overcome this drawback,two new redesigned networks,namely,backbone network and additional networks are proposed.Their motives are the same as those of deep and shallow networks.This paper also studies the deployment of the network,and the deployment prototype verification system is developed.In this paper,the parameters of the backbone network and the additional network are compressed to 55628 and 7930,which are much less than those of a similar modulation recognition network(most parameters range from hundreds of thousands to millions)and combination network(about one hundred thousand).Then,the model is transformed into the ONNX model,and the prototype verification system is developed by C#.The simulation result on the public dataset shows that the lightweight combined network achieves a high accuracy of 98.9%,slightly 0.2% higher than that of the combined network.Also,from the simulation results,it is evident that the proposed design greatly reduces the number of parameters,which demonstrates its effectiveness.Simultaneously,using the prototype verification system of the transformed model,under the premise of a small batch size,the average single sample predicting speed is higher than that of the model before the transformation,which proves the feasibility and efficiency of the deployment.