Research On Data Cleaning Of Electron Cyclotron Emission Imaging Based On Machine Learning

In diagnostics on tokamak,Electron Cyclotron Emission Imaging(ECEI)plays an important role in measuring the electron temperature profile.Hence,diagnostic data can be used to study sawtooth,which is vital to understand sawtooth precursor mode.In order to dig out statistical rules of diagnostic data more conveniently,it is necessary to perform data cleaning on the ECEI signals in the early stage including filtering good signals and making a preliminary classification of the signals.ECEI signals can be classified into categories of saturated signals,zero signals,weak signals,and normal signals preliminarily.Traditionally,the ECEI signals are classified by manual approaches,but remaining a problem of low recognition efficiency and large error rate.In order to achieve the goal of the classification of the ECEI signals,machine learning methods are used to address the duty of classifying the ECEI signals based on the big data of the raw ECEI signals.In view of the classification of the ECEI signals is a multi-classification problem,the specific classification algorithms from two angles are realized according to the strategy of machine learning.One is to disassemble the multi-classification problem into multiple two-class problems,the combination of support vector machine(SVM)and decision trees(DTs)is used to gradually complete multiple two-class classification tasks.SVM and DTs are both high-efficient algorithms which have good performance on two-class classification tasks.The second one is to use machine learning algorithms that can directly deal with multi-classification problems,such as random forest(RF)in ensemble learning.In order to compare these machine learning methods,the combination model of SVM and DTs,and RF model are achieved to complete the task of classifying the ECEI signals on the same data sets.The effect of overfitting is further eliminated by using a five-fold cross-validation approach when training the model on the dataset.From the final experimental results,two models have their own advantages and disadvantages.In the data set,the combination model of SVM and DTs is stronger than the RF model in the recall rate of weak signals.However,in the test set,the generalization error rate of the combination model is higher than the RF model.The application of these two machine learning models is a preliminary study in the big data research of the ECEI signals.They show quite good classification performance on small validation sets,and two models have their own advantages and disadvantages.In order to strengthen the comparison between two models,the performance evaluation on a larger data set is required.This paper introduces machine learning into the classification of the ECEI signals,which accelerates the data cleaning of a large amount of the stored experimental data.The feasibility of using machine learning to process scientific big data is also verified.Besides,machine learning methods can be applied to extract relevant physical patterns from the huge amount of experimental data in the future research.It is obvious that machine learning plays a more and more important role in the study of tokamak experimental data.