The Intelligent Method Of Filtering Pulsars Based On Survey Data

At present,there are many radio devices invested in pulsar survey observation,which have generated more and more candidate images,and most of the candidate images are background noise or useless information.Traditional artificial recognition methods have the disadvantages of low efficiency and high subjectivity when filtering pulsar candidates from the candidate images,the training of existing machine learning recognition models depends on the artificial design characteristics.The purpose of this thesis is to build a high-efficiency pulsar candidate intelligent filtering model which use the candidate images as the training samples of a model to achieve direct recognition and classification of the candidate images.Support Vector Machine?SVM?has the advantages of clear mathematical explanation,no local minimum value problem and strong generalization ability.Compared to fully connected neural network,Convolutional Neural Network?CNN?have the advantage of automatically extracting useful information from images,a fewer parameters and lower complexity.SVM and CNN have their own advantages in image classification tasks.SVM has strong generalization performance,and CNN has the advantage of learning potential commonality of samples.In this paper,they are applied to pulsar candidate filtering respectively.The specific research contents are as follows:?1?Dataset construction.In this paper,only the measured pulse contour images?grayscale?obtained from the CSIRO Observatory are used as samples,which contains 21 pulsar observation images,including single-peak,bimodal pulsar candidate images?positive class?and non-pulse candidate images?negative class?,which are referred to herein as original images.A training set CT,a testing set CV,and a total set CZ compose of the original images.A training set ST,a testing set SV,and a total set SZ compose of one-dimensional feature vectors which are obtained from the original images by single-channel graying and one-dimensional expansion.It should be mentioned that the total sets contains the samples of the training set and the testing set.?2?An SVM method for adapting pulsar candidate features is proposed..Scikit-learn is used to build the S₀ model,and the ST and SV are respectively used to train and evaluate the model.Because the pulse contour images contain a lot of redundant information,in order to find a model with lower complexity and better performance,S₁^S₁3 models are constructed.the feature vector used to train and evaluate models are obtained by different dimensional PCA dimensionality reduction of the ST and the SV.The experimental results show that: 1)When the feature vector is reduced to not less than 1200 dimensions,the spatial distribution of positive and negative examples is linearly separable in the training set;2)In a certain dimension reduction range,the PCA dimensionality reduction can make the model's recognition rate of the positive class remain stable in the testing set,but the recognition rate on the negative class will decrease.The reason for this result may be that the negative class images itself are mainly generated by noise folding,and the purpose of PCA dimensionality reduction is to remove noise;3)S₀ has the best performance,and the recognition accuracy of the positive and negative examples is 96.4% and 92.8% in the testing set respectively.Because the negative samples of the data are not sufficient,the constructed training set is a small sample size,so the recognition rate of the negative samples is lower than that of the positive samples in the test set.?3?The CNN-based pulsar candidate filtering method is studied,and a CNN structure for pulsar candidate recognition is proposed.In order to find a relatively optimal model,Tensorflow is used to construct five different CNN models LC₁^LC₅.The LC₃ model is used to analyze the influence of different learning rates and different optimization algorithms on the performance of the model.The experimental results show that the performance of the LC₄ model,which consists of one convolution layer and two hidden layers,is best,and the accuracy on the testing set CV achieves 99.8%,when the learning rate is 0.001 and the optimization algorithm is Adam.?4?In order to evaluate the generalization performance of different algorithms,the trained LC₃,LC₄ and S₀ were tested on the total set respectively.The experimental results show that the recall rates?false positive rates?of LC₄,LC₃,and S₀ on positive and negative examples are 98.99%?2.39%?,98.4%?3.86?,and 97.34%?2.33%?respectively.The above results show that CNN's model performance is better than SVM,but the performance of the SVM is acceptable.It should be mentioned that training CNN is difficult and time consuming,in contrast,SVM is easier to train.In actual use,LC₃ is used when a high classification requirement is required for positive samples,and S₀ can be used when roughly filtering candidate samples.In the field of pulsar candidate intelligent search,compared to use multi-feature to train models in foreign research institute,the CNN model in this paper adopts only pulse contour images,which eliminates the cumbersome feature design of the previous period.But the generalization performance,especially in filtering pulsar candidate samples,is excellent.Compared with the domestic algorithm,the methods proposed in this paper achieve better recognition results with less training samples,and augment the indicators of false positive rate and recall rate,which can evaluate the performance of the model more reliably.