Research On MFR Working Pattern Recognition Based On Machine Learning

The research of radar work pattern recognition is one of the current research hotspots of cognitive electronic warfare.At present,many mainstream classification methods of machine learning are slightly insufficient in the behavior recognition of MFR(Multi-Function radar,MFR),which will cause problems such as low generalization ability and inaccurate classification.At the same time,the imbalance data also add a certain degree of difficulty to the behavior identification of MFR in practical applications.Therefore,this paper aims at accurately and efficiently classifying MFR working modes.And further research is carried out in view of the problems of current classification methods in the complex parameter features and unbalanced samples.First,SVM(support vector machine,SVM)is a model that is good at classifying big data,and it is suitable for the balanced samples of the MFR working mode in this article.combined with the working principle of SVM,the influence of parameter values on the classification effect of SVM under balanced data is analyzed.The four SVM parameter optimization algorithm grid search methods,PSO(Particle swarm optimization,PSO),PSO improved algorithm and adaptive simulated annealing algorithm(Simulated Annealing Algorithm,SA)are introduced in detail.The simulation results show that the classification ability of SVM with fixed parameters is not as good as the one with parameter optimization.PSO and adaptive optimization algorithms have advantages in optimization time and accuracy respectively.Second,based on the parameter optimization of SVM,research is carried out on the problem of attribute overlap in MFR working mode data when classifying balanced data.To solve this problem,an SVM-NP classification model based on SVM-RFE algorithm is proposed.The model is based on the SVM-PSO,and composed of K-NN boundary sample pre-selection and SVM-BP algorithm to achieve double reduction of boundary weight reduction and classification weight reduction.Simulations and comparative experiments prove that the SVMNP model has a better classification ability in the recognition of MFR working patterns of balanced samples.And it effectively reduces the attribute overlap of each pattern parameter to a certain extent.Third,in the perspective of data sampling,the MFR work pattern recognition of unbalanced samples is deeply studied.To improve the data loss and sample redundancy caused by the classic unbalanced sample sampling algorithm,this paper improves the Smote(Synthetic Minority Oversampling Technique,Smote)algorithm to obtain the Weighted-Smote oversampling algorithm.Weighted-Smote reduces sample redundancy by assigning higher weights to samples near the border and center when synthesizing samples,while also avoiding data loss.After simulation and comparison,it is concluded that in the MFR work pattern recognition of unbalanced samples,the correct classification of samples after sampling is higher.Among them,the Weighted-Smote algorithm has certain advantages in the recognition of samples of various unbalanced degrees.Finally,the integrated algorithm of MFR work pattern recognition under unbalanced data is studied.To solve the problem that random forest is restricted by poor classification,the voting random forest algorithm combined with oversampling algorithm is proposed.Aiming at the shortcomings of the Boosting structure,a residual optimization structure based on Boosting is proposed.Through simulation and comparison,the two proposed methods make up for the shortcomings of the current integration algorithm,especially the classification accuracy of the residual optimization structure is significantly improved.