Research On Multi-signal Modulation Type Recognition Method Based On Deep Learning

Signal modulation recognition is widely used in electronic reconnaissance equipment,which is of great significance and practical value to improve the technical ability and level of electronic warfare.In the current information battlefield,the number of various radar,communication and other radiation sources is increasing.Modulation forms are becoming more and more diverse.The increasing signal density makes the electromagnetic environment more and more complex.How to identify multi-component signals accurately and quickly from such a complex,dense and changeable electromagnetic environment is an urgent problem to be solved.Aiming at the problems of low accuracy,poor robustness and poor real-time performance of existing modulation recognition methods for multi-component signal recognition in complex electromagnetic environment,this paper puts forward corresponding solutions.The main research contents of this paper are as follows:The types of radar intra-pulse modulation signals used in this paper are introduced,the fast Fourier transform and time-frequency analysis method are used for simulation analysis,and the characteristics of the two processing methods are summarized.Finally,the multi-component signal data set is constructed,including spectrum data set and time-frequency image data set.The methods of data generation and data calibration are developed respectively,which lays a foundation for the implementation of subsequent multi-component signal recognition algorithms.Aiming at the problems of poor real-time and low accuracy of existing multi-component signal recognition methods,a multi-component signal separation and recognition algorithm based on one-dimensional detection network is proposed.Firstly,the spectrum data of the signal is obtained,and the feature extraction layer based on one-dimensional convolutional neural network,cyclic neural network and complex neural network is built.On this basis,the onedimensional detection network,loss function and corresponding test methods are designed.The separation,location and recognition of time-domain aliasing and frequency-domain partial aliasing signals are realized.Compared with the traditional frequency domain filtering method,the one-step detection greatly reduces the algorithm steps and speeds up the recognition speed.Aiming at the problem of serious time-frequency domain aliasing and poor robustness of multi-component signals,a multi-component signal separation and recognition algorithm based on semantic segmentation is proposed.Firstly,the improved Cohen time-frequency distribution and convolution denoising autoencoder are used to filter the signal in two steps.Full convolution network,U-Net and Deep Lab V3+ are used for training.A suitable loss function is designed to achieve positive and negative sample equalization.Finally,a color threshold filter is built to correct the segmentation results,and the separated single signal binarization timefrequency map is obtained.The algorithm realized the separation and recognition of fourcomponent time-frequency domain aliasing signals.