Research On Pattern Recognition Method Based On Compressive Sensing And Random Forest

With the development of the information,people deal with a large number of complex information every day.Pattern recognition can automatically divide a large amount of information into their own pattern classes by using computers,and it has been widely used in more and more fields.The purpose of this paper is to provide an idea of constructing a pattern recognition system model.By using feature extraction and classifier selection,the process and selection criteria of model building are discussed.Finally,a framework suitable for any pattern recognition is given.In view of the above,this paper carried out the following work:First of all,three traditional dimensionality reduction methods for principal component analysis and compressive sensing are analyzed and compared in this paper.And these principles and application in feature extraction of dimensionality reduction are studied in depth.In addition,this paper introduced several common classifiers and described the random forest classifier in detail.Secondly,a new robust dimensionality reduction method is proposed,which can remove some noise after reducing the dimension and improve the classification accuracy.The validity of the proposed dimensionality reduction method is verified from three aspects: energy proportion,intra-class distance and inter-class distance between samples,and the recognition accuracy after dimensionality reduction.At last,a modulation recognition method based on dimension reduction is designed,the system model is for 2FSK,BPSK,QPS,MSK and 2ASK modulation signals.The recognition rate of this five kinds of modulation signal is greater than 90% when SNR is greater than-10 dB.Finally,the voting process of random forest classifier is improved properly.An improved random forest algorithm is proposed to improve the classification performance under noisy environment.Furthermore,a communication radio recognition based on transient signal is designed.Ten Motorola interphones are taken as the research object,and the transient signals are collected without considering the influence of multipath,time delay,temperature and equipment aging.High-performance Agilent oscilloscopes are connected to communication stations using cables,and power-on transient signals are collected directly at a sampling rate of 40 MHz.The number of sampling points is 159,901.In the case of Hilbert transform envelope characteristics of the transient signal for the initial feature,three principal components method was compared by analyzing the energy retention,intra class distance and classification accuracy of the sample after dimensionality reduction,and we obtained the best dimension reduction method.In different dimension samples as the input condition,random forests,support vector machine,BP neural network,and grey correlation analysis was compared to recognize communication radio.And using DS evidence theory,the random forest classifier was improved to obtain higher classification accuracy.At last,we got a complete pattern recognition system model.The improved random forest classifier has a high recognition rate of 90% for the communication radio in this experiment under the condition that the SNR is greater than 6dB.