Research On Link Adaptive Decision Technology Based On Artificial Intelligence

As the growth of communication technology,derived from the perception of electromagnetic environment and optimizing transmission decision,it becomes a trend to obtain a better compromise between transmission performance and efficiency.Especially in the complex electromagnetic environment,while ensuring the robustness of the communication network,it can support the free configuration of the communication link following the real-time situation.Based on the compromise between performance and complexity,the classical forward continuous mean elimination algorithm is selected for interference detection based on energy difference.Combined with the time-frequency characteristics of different typical interferences,a time-frequency domain cascade interference detection scheme is designed.The digital simulation results show that the jamming detection rate can reach more than 90% when the jamming to noise ratio reaches 5d B.The thesis studies the estimation technology of typical interference parameters.Firstly,the thesis deeply analyzes the signal and noise power estimation technology based on symbol splitting moment,second-order periodic graph,second-order and fourth-order moment and maximum likelihood.The symbol splitting moment and the second-order periodic graph method do not need priori signal,and have a wide range of application but limited performance.The second-order and fourth-order moment method has low complexity but poor performance.Therefore,the maximum likelihood estimation algorithm based on known training sequence is selected in this thesis.The thesis further designs the maximum likelihood estimation scheme of joint signal power,noise power and interference power under the condition of interference.The digital simulation conclusions prove in the slow fading condition,the deviation of estimated signal-to-noise ratio is no more than 1d B,and the mean square error of interference power estimation is no more than one percent.Analyzing and comparing the typical characteristics of interference,this thesis selects the normalized 3d B bandwidth,frequency domain moment kurtosis and normalized spectrum kurtosis coefficient with higher discrimination as the interference characteristics.This thesis analyzes and models the mainstream machine learning methods represented by backward propagation neural network,support vector machine and decision tree.The digital simulation results show that with the reduction of the jamming noise to ratio,the artificially selected interference characteristics are gradually blurred,and the recognition effect of the three methods is not ideal.Even if the methods are combined to improve the recognition effect to a certain extent,it still can't meet the system requirements.The thesis further studies the interference recognition method based on convolutional neural network(CNN),selects the time-frequency image with more abundant information as the / training input,and the CNN automatically extracts the features;Through analysis and comparison,the thesis uses the Choi-Williams distribution with richer and more accurate image details to replace the traditional short-time Fourier transform,and improves the defect of multi-tone time-frequency image deformation through cross term interference suppression.The thesis further optimizes the design of convolution layer and convolution kernel by combining the complexity of parameter space and recognition performance.The simulation results demonstrate that the total recognition accuracy can be improved by 7% under the condition of traditional CNN learning ratio.Aiming at the conditions that the target interference samples are difficult to obtain or there are differences between the source interference and the target interference in parameter,this thesis studies the interference recognition problem in the few shot situation.Based on the applicability of the algorithm and the test results on the standard few shot data set,a meta learning algorithm independent of the model is selected to improve the interference recognition performance in the few shot scene combined with data expansion.This thesis describes the complete process of task sampling,meta training and meta testing of support set and query set.The digital simulation results show that the interference recognition rate can still be maintained at more than 90% under the condition that there are differences between the source interference and target interference in parameters and the target interference can obtain few samples.On the premise of ensuring the reliability of communication,the thesis builds an adaptive decision system modelled on decision tree with the interference parameters,categories,signal-to-noise ratio environment and the mapped optimal link as the input.This thesis compares the simulation performance of traditional communication and intelligent decision-making based on machine learning.It can be seen that on the communication BER constrain,intelligent decision-making method can effectively improve the transmission rate.Thesis constructs a system level simulation platform and develops a visual interface,which has the function of real-time observation of interference and signal spectrum situation,and shows the function of parameter estimation,interference detection and identification results and link decision-making.After testing,it meets the expected design function.