Intelligent Detection And Parameter Estimation Of Electromagnetic Signals Based On Mask R-CNN

In today's rapid development of life,the signal detection and parameter estimation technology of communication signals has always been a hot research topic.With the development and advancement of science and technology,the types and combinations of received signals are intricate,and multiple signals with overlapping time and frequency domains account for the majority.Therefore,how to detect and estimate the parameters of overlapping multiple signals in the timefrequency domain becomes the first problem to be solved.Traditional signal detection and parameter estimation methods are difficult to effectively perform visual detection and intelligent parameter estimation.In order to solve the signal detection and parameter estimation problems of overlapping multiple signals in the time-frequency domain,this paper combines artificial intelligence,and launches the research of intelligent visual detection and intelligent parameter estimation methods for overlapping multiple signals in the time-frequency domain.This paper analyzes the basic knowledge of neural networks and common convolutional neural network models and recurrent neural network models.On this basis,a time-frequency domain overlapping multi-signal intelligent detection method based on masked area convolutional neural network and LSTM network are proposed The main work of this dissertation are as follows:1.The current research of traditional signal detection and parameter estimation and the research status of signal detection and parameter estimation based on deep learning are systematically explained.The basic knowledge of deep learning network is introduced,and the structure of regional convolutional neural network,fast regional convolutional neural network and masked regional convolutional neural network are introduced in detail,which lays the foundation for the intelligent detection of overlapping multiple signals in the time-frequency domain.Also,the structure of cyclic neural network and long-short-term memory artificial neural network is introduced to lay the foundation for intelligent parameter estimation of overlapping multiple signals in time-frequency domain.2.An intelligent detection algorithm for overlapping multiple signals in time and frequency domain based on masked area convolutional neural network is proposed.Firstly,the short-time Fourier transform for overlapping multiple signals is performed to obtain the time-frequency spectrum.Then,the Labelme is applied to do data preprocessing on the time-frequency spectrum to generate the corresponding target signal area and the corresponding mask area image.Next,all the obtained data simultaneously input to the Mask R-CNN network.At the same time,to solve the problem of missing overlapping position information of multiple overlapping signals in the timefrequency domain of Mask R-CNN network,an improved fusion algorithm,Criminisi algorithm is proposed.Before the Mask R-CNN network extracts the target pixel information,the Criminisi algorithm is used to repair the location pixel information.The simulation results show that the improved algorithm has doubled the training speed of the original algorithm,and the recognition performance has increased by 3d B on average.3.A parameter estimation algorithm of linear frequency modulation signal based on LSTM network is proposed.First of all,according to the detection results of the above algorithm,the initial frequency and FM slope parameters can be estimated from the position of the signal pixel edge point,and the estimated initial frequency and FM slope can be used as additional input information for the LSTM network.Then,blind source separation is performed on the chirp signals overlapping in the time-frequency domain to obtain a single signal,and the single signal is subjected to STFT to obtain a time-frequency diagram.Together with the initial frequency and FM slope parameter information obtained by the above algorithm,they are used as the input of the LSTM network.The initial frequency and FM slope are estimated.Finally,a single signal is used as the input of the LSTM network to estimate the amplitude.The experimental results show that the relative error rate of the parameter estimation of LFM signal based on the LSTM network is 5%,which is 6% lower on average than the method based on the detection results,and is 23% lower on average than the method based on FLOSTFT-Hough.