Design Of Efficient Neural Network For Radar Signal Recognition

The technology of radar signal recognition is widely used in military and civilian fields,especially in military scenarios,it plays an important role in modern electronic countermeasures and is a very challenging problem.With the rapid development of deep learning technology,radar signal recognition technology based on deep learning has also made great progress.Compared with traditional recognition methods,the overall performance has been greatly improved.While the recognition accuracy continues to improve,the depth of the neural network continues to deepen,and the number of parameters continues to be redundant.The requirements for machine computing power are also very large,which makes it difficult to apply on edge terminal devices.It is very important to design efficient and streamlined neural networks.In this paper,a radar signal recognition method based on neural network architecture search is proposed.Compared with radar signal recognition based on deep learning,in terms of network design,this method automatically designs and forms a convolutional neural network through a pre-designed search space and search strategy.There is no need for relevant researchers to spend a lot of time and energy on the design of the network,at the same time,it also have certain advantages in the performance of the network.The simulation results show that the radar signal recognition method based on neural network architecture search has high classification and recognition accuracy and consistency,and also has a good performance in the complexity of the network model.Compared with traditional convolutional neural networks,the number of parameters and floating-point operands are greatly improved.This method achieves an average recognition accuracy of 93.2%with 1.28GFLOPs of floating-point operations and 2.57×10~6parameters,which has obvious advantages over other traditional methods.Considering that the network search phase in the neural network architecture search method is relatively time-consuming,this paper adopts the lightweight method of network pruning in deep learning.Simulation experiments on traditional convolutional network models and lightweight neural networks show that network pruning has a certain good effect in reducing the complexity of the model,but it needs to sacrifice part of the recognition accuracy of the network.At the same time,the effect of this method is more obvious on the network model with more redundant parameters.Compared with the neural network architecture search method,it does not need to build a new network separately,but there is a small gap in the recognition accuracy of the network.