Fault Identification Location And Fault Tolerant Control For Industry System Using Tensor Label Learning

The core of supervisory learning is that the data has labels.When labels are used to describe fault classes location and fault tolerant control,there are great limitations for the existing label methods.Based on the the industrial process characteristics,data characteristics and machine learning fault diagnosis methods in the production process,the fault identification,location and control of faults are studied in this thesis by tensor label learning strategy.Heterogeneous information output from the three fault diagnosis parts are unified into the framework of the tensor learning.This thesis studies the problems of identification,location and fault tolerant control,and realizes the unified representation of heterogeneous information data in the problems of identification,location and fault-tolerant control.The unified representation and learning performance evaluation of heterogeneous information data is realized by means of the method of unified representation of the information data of the high-dimension information data in the process of industrial fault diagnosis.The identification,location and control of faults are studied by using the method of tensor label learning.The different information is fused into the framework of tensor learning,and different information is fused into the fault diagnosis tensor,and the fault processing capability of the fault diagnosis system is comprehensively evaluated.The main contents of this thesis are as follows:Firstly,the framework of tensor label learning is established.In this section,the concept of tensor label learning is proposed,and the supervised learning is extended to high dimensional tensor form.This study integrates supervised learning and multi-label problems into the framework of tensor label learning as a special case of 0-order tensor and 1-order tensor learning.In order to realize the uniform representation of the heterogeneous information data existing in the problems of identification,location and control,the 0-order tensor label learning(scalar label learning),1-order tensor learning(vector label learning)introduced in the previous chapters is extended to the form of the matrix label learning and the third order label,and the evaluation index of the corresponding learning performance evaluation is derived.Secondly,the problem of simultaneous fault identification is studied in this paper.Because of the combination of single failure modes,the problem of simultaneous faults is difficult to identify.A method of vector label learning is proposed to solve the problem of simultaneous faults.Only independent n-type single-fault training data can be used to predict the type of 2n concurrent faults.Thirdly,aiming at the problem of scarcity of simultaneous fault data,this section presents a vector label-transfer learning method for solving the problem of scarce training data of simultaneous faults in industrial systems.The simultaneous fault recognition system includes two parts: vector label-transfer learning framework and random forest machine learning algorithm.The source domain(single fault data)and target domain(concurrent fault data)of this study are different distributions,and the source task and the target task are the same,and the strategy to encode the quantization label links the two data domains with different distribution.The experimental results show that the method can quickly obtain the model of simultaneous fault identification using very few target domain data.Forthly,in order to solve the problem of fault location based on data mode,the condition of fault location based on feature selection is proposed.Then,the fault location system based on similarity measure is studied,and a new feature selection algorithm is adopted,which is based on similarity measurement between different types of fault samples.At the same time,the correlation between the features is removed to select the features most relevant to the fault type to achieve fault location.In this study,the optimal Relief F feature selection algorithm is used to select the most relevant features of the fault category,and the location of the model feature collection is used to locate the fault,and the feasibility of the method is verified by the experimental simulation on the SOFC system.In order to solve the problem of fault identification and location simultaneously by using a set of algorithms,this chapter adopts the method of random forest used in the previous chapter,and selects a suitable subset through the characteristic measure of random forest,so as to determine the location of fault.Qualitative analysis is used in fault location in this chapter.Fifthly,in order to solve the problem of the aging failure,this chapter adopts the same type of fault and collects the data from different working conditions to identify the fault classes,and studies the invariant characteristics of the fault type by the method of supervised learning which is also regarded as 0-order tensor learning.Thus the fault information mining is realized.Based on the results of fault identification,an active fault tolerant controller with control law rescheduling based on the result of fault identification is designed in this section to realize the fault recovery of industrial system.