Research On Mutual Information Feature Selection Algorithm Based On Weight Estimation And Fault Diagnosis Of Nuclear System

Due to the large amount of data collected by nuclear power simulation system and the increasing development of deep learning,the data-driven artificial intelligence fault diagnosis technology is getting more and more attention.And in the engineering practical problems,people prefer to use the fault sensitive feature set for fault diagnosis,the selected sensitive feature set is more conducive to fault location and analysis in system operation and maintenance,and it is more immediate to repair and sort out the system when the sudden change of fault conditions occurs,so this paper carries out the selection of fault critical features before fault diagnosis.In this paper,we propose a feature selection algorithm based on conditional mutual information weight coefficients with minimization(FWCM)for the problem of incomplete removal of redundancy in key feature sets in the joint mutual information(JMI)algorithm for thermal hydraulic systems of nuclear power plants.Due to the fact that existing algorithms do not consider the different effects of selected features on the correlation of candidate features,in the process of dynamic feature selection,this paper derives the weight estimation formula w as a measure of the importance of different selected features to new classification information.At the same time,considering the high redundancy of nuclear power data and a few features,the minimum residual classification information is used to replace the average residual classification information,To reduce the problem of overestimating residual correlation previously,the feasibility and effectiveness of the proposed algorithm were verified on 8datasets of UCI.In order to evaluate the effectiveness of different classification accuracy obtained from different feature sets selected in nuclear power system experiments,this article proposes an evaluation index(EI)as the basis for selecting key features,in order to more accurately and objectively select key feature sets.Different subsets selected by FWCM algorithm and JMI algorithm are used in 1D-CNN-LSTM,2D-CNN models for fault diagnosis,and the proposed EI metrics are used to select the key feature sets,and finally the feasibility analysis of the algorithm is carried out based on the diagnosis results.