Research On Satellite Communication Signal Parameter Estimation And Interference Situation Analysis Technology Based On Deep Learning

With the rapid development of low-orbit satellite constellation,especially giant loworbit satellite constellation,the types and quantities of frequency equipment are increasing day by day,and the available spectrum of satellite communication system is limited,so it is very necessary to estimate the parameters of satellite communication signals.Among them,signal sign rate estimation and interference detection can be used to assist spectrum sharing between satellites.However,in the face of the large time delay between space and earth,only sensing the current interference characteristics can no longer meet the needs of efficient control.Therefore,in order to achieve the safe control of spectrum situation,the following problems will be faced: 1)How to improve the accuracy of signal symbol rate detection;2)How to improve the robustness of detection of burst interference;3)How to improve the prediction accuracy of interference occupancy power spectral density value based on historical perception data.Compared with traditional methods,artificial intelligence technology is becoming more and more mature and is widely used in various fields of daily life.Therefore,in order to improve the processing efficiency and capability of satellite spectrum data,this article studies symbol rate estimation method,semi-supervised learning interference detection method and interference occupancy state prediction method based on deep learning.The main research work and innovation points are as follows:Firstly,Aiming at the problem that the estimation accuracy of the traditional cyclic stationary symbol rate estimation method is not high when the SNR is low,a convolutional neural network based symbol rate estimation method is proposed.The proposed method mainly includes three stages: data preprocessing,offline training,and online testing.Specifically,the feature vectors of perception signals are extracted based on cyclostationarity theory,along with their corresponding symbol rates as labels,and fed into a neural network for training.The trained model then predicts the symbol rates based on the feature vectors.Experimental results show that compared to methods based on cyclostationarity theory,the proposed approach achieves higher detection probability.Secondly,Aiming at the low robustness of Yolo model based interference detection algorithm,a semi-supervised learning based interference detection algorithm is proposed.It consists of three parts: dataset creation,network design,and model training.To address burst interference,a Faster R-CNN network is introduced to construct a teacher-student model,which identifies and classifies interference with only a few labeled images.Experimental results demonstrate that compared to traditional supervised detection models,the selected method achieves an m AP of 79.37%,significantly improving the robustness of interference detection.Finally,In order to solve the problem of low performance of traditional long short-term memory neural network in the space to Earth propagation environment,a new algorithm combining long short-term memory neural network and autoregressive moving average model is proposed.The proposed method mainly includes three stages: spectrum data preprocessing,offline training,and online testing.The network parameters were optimized by redesigning the loss function of the model training and combining the backpropagation algorithm to improve the prediction accuracy of spectrum occupancy status.In addition,considering the correlation between the occupancy status of the same interference at different frequencies,a fusion network was designed to further improve the performance of the proposed prediction method.Experimental results show that compared with traditional LSTM,ARMA,and CNN-Bi LSTM models,the proposed method can achieve the highest occupancy status prediction accuracy.