Data-driven Fault Detection Technology And Application

Data-driven fault detection is more suitable for diagnosing complexity spacecraft rather than other methods.Power system is one of the complexity systems in satellite,which is also the significant part to ensure the normal operation of the satellite.In this paper,data-driven fault detection technology is studied by using satellite power system.Data-driven fault detection is regarded as multiple attribute decision making issue including the generation of decision label,attribute subset selection,decision making,and the challenging of massive data and new-coming data,which are studied successively.The main researches are shown as follows:Labeled multiple attribute decision making is raised in terms of the characteristic of data-driven fault detection,which is the basic theory of the following research.Firstly,according to the data systems(information system and decision system),the processing of data mining and decision making in data-driven fault detection,and combining with the theory of multiple attribute decision making,the basic theory of labeled multiple attribute decision making is proposed.Secondly,the goal of labeled multiple attribute decision making is raised by using the principle of decision risk minimum and the purpose of fault detection.Finally,the essence and existing problems of labeled multiple attribute decision making are deeply analyzed,and the requirements and assumptions for subsequent are put forward.In order to solve the problem of knowledge acquisition in information system,the neighborhood grid clustering and the neighborhood density grid clustering methods are proposed to realize the transformation of information system into the decision system.Firstly,aiming at the shortcomings of the existing clustering methods,a clustering algorithm based on grid subspace partition is raised.The clustering experiments based on artificial data show that the proposed method does not need to set the number of clusters in advance and searches for the cluster centers independently,meanwhile it can deal with the clustering of data with arbitrary shape.Then,in order to deal with overlapped data,the neighborhood density grid clustering method is proposed by considering the local density.Numerical experiments show that the neighborhood density grid clustering method inherits the advantages of neighborhood grid clustering,moreover it can effectively deal with overlapped data.Subsequently,the proposed theory is applied to typical fault detection of satellite power system.Experimental results show that the proposed clustering method is better than other common clustering methods.Finally,some discussions and parameter robustness analysis of the proposed clustering method are given.From the perspective of statistics,a fuzzy Bayes risk model is put forward to achieve knowledge acquisition and decision making in decision system.Firstly,from the distribution characteristics of the data,the potential information of the data is deeply analyzed,and the fuzzy Bayesian risk model is proposed in combination with the target of labeled multiple attribute decision making.Then,the model is applied to the selection of attribute subset and attribute weighting in the label multiple attribute decision making,and the experimental results show that the proposed theory and method have a certain advantage.With regard to the evaluation of attribute weighting,the correlation coefficient with the longitudinal deviation and transverse residual is proposed,and the results of artificial data show that the proposed index is more reasonable.Then,fuzzy Bayes risk model is combined with T-S fuzzy model to achieve rule extraction and decision making.Finally,the proposed method is applied to the typical fault detection of satellite power supply system.The experimental results show that the proposed method is suitable for the fault detection of satellite power system,meanwhile the superiority and rationality of the method is verified again.In the final decision experiment,the hypothesis of labeled multiple attribute decision making is verified.In order to make up the shortage of the statistical idea,the single parameter decisiontheoretic rough set model and the neighborhood single parameter decision-theoretic rough set model are put forward to realize the knowledge acquisition in decision system.Firstly,the shortage of knowledge acquisition and decision making based on the statistical idea is analyzed,the data information is depicted in rough approximation space,and the theory of decision-theoretic rough set and three-way decision idea are deeply analyzed.In order to solve the problems of too many parameters and parameter uncertainty in the decision-theoretic rough set model,a new parameter determination formula is derived,and a single-parameter decision-theoretic rough set model is put forward with the corresponding parameters analyzing and discussion.Then,the neighborhood relation is used to instantiate the single-parameter decision-theoretic rough set model,and neighborhood single-parameter decision-theoretic rough set model is therefore presented,and the attribute reduction and weighting algorithms are designed.In the experiments,the single parameter selection principle is given,and the attribute reduction and weight assignment are compared.The results show that the single-parameter decision-theoretic rough set model is more practical than the traditional decision-theoretic rough set model.Finally,the practical application of the proposed theory is verified in the fault detection of satellite power system.In view of the problem of fast mining in massive data and the influence of new-coming data on knowledge base,a grid-clustered rough set model and rough self-learning theory are proposed to implement fast attribute reduction and the function of self-learning on new-coming data,respectively.Firstly,combined the clustering idea with the rough theory,the grid space clustering algorithm and grid-clustered rough set model are proposed,and the corresponding heuristic reduction algorithm is designed.Then,based on the idea of knowledge partitioning,rough self-learning theory is proposed and instantiated by the grid-clustered rough set model.In the numerical experiment,the test of parameters selection,the comparison experiments on clustering method selection,attribute reduction result and reduction speed,and the verification experiment of rough self-learning are completed in turn.The results show that the proposed theory and method have great advantages in fast attribute reduction,and can effectively solve the impact of new data on knowledge base.The study of the above problems belongs to the advanced technology of detection.