Research On Radar Signal Recognition Based On Automatic Machine Learning

Today,the modern war is changing to the direction of information.In the wider and wider range,electronic warfare is affecting the development of war.As the forerunner and foundation of electronic warfare,electronic reconnaissance has become an important factor in deciding the outcome of war.In this context,the recognition of radar emitter signals has been studied in this paper,and in the process of solving the hyperparameters optimization problem of recognition model,I have further studied the automatic machine learning methods for algorithm selection and hyperparameters optimization in the frontier field.This new method of automatic machine learning provides a new idea for pattern recognition,which is very suitable for radar signal recognition.After extracting the features,this paper researched the recognition of radar signals based on AUTO-SKLEARN and TPOT.First of all,based on the research on feature extraction of radar emitter signal by the tutor research team and the radar intelligent perception system developed by the author in the related research project,aiming at the emitter signals in research projects,this paper extracted the information entropy,wavelet ridge frequency feature,and the two complexity features of the information fractal and Lempel-Ziv complexity.Besides,in order to provide more useful features for automatic machine learning,we used the Denoising Autoencoder of deep learning model to extract features of radar emitter signals.Through a complex network and efficient training methods,deep learning can find deep features which can not be found by general methods.And By adding noise to the original signal,Denoising Autoencoder can discover and express the essential feature of the original signal in the process of coding and reconstruction.Then in order to solve the CASH problem of algorithm selection and hyperparameter optimization in radar signal recognition,this paper used AUTO-SKLEARN to recognize the extracted features.In order to solve the CASH problem,AUTO-SKLEARN uses a Bias optimization tool named SMAC which is based on the random forests to realize the algorithm parameter optimization.AUTO-SKLEARN makes a hot start for Bayesian optimization by meta learning,which specifies the candidate scheme for optimization,and reduces the search space.Furthermore,it makes up a number of models with better evaluation results as an integrated model,which can maximize the effectiveness and avoid overfitting as much as possible.Besides,this paper also adopted TPOT,which can generate any tree structure pipeline.The algorithm of every node in the pipeline can be randomly matched,and it is evolved by genetic programming to optimize the most suitable machine learning pipeline.The experimental results show that the recognition performance of AUTO-SKLEARN and TPOT is very good and stable.Finally,through the comparative analysis,using TPOT to process features alone,SVM achieves the best recognition effect,and AUTO-SKLEARN can play the comprehensive advantages of multi classifier.Therefore,this paper proposes to process features with TPOT,and then recognize the transformed features with AUTO-SKLEARN.By combining these two automatic machine learning methods,the recognition effect is further improved.