| It is complex for chemical process with high nonlinear,continuous and time-varying characteristics.Once the failure occurs,it will bring serious loss to the economy and life safety.Therefore,early warning plays a very important role in chemical process monitoring.How to excavate useful information from online massive industrial data and carry out fault diagnosis of chemical process has become a hot research topic at present.With the development of artificial intelligence,fault diagnosis technology has entered a new era.However,the problems of fault diagnosis with incomplete fault information and multiple fault diagnosis need to be further explored.Rough set theory and fuzzy set theory are two important tools to deal with incomplete and uncertain data in the field of artificial intelligence.At present,fuzzy set theory has been widely used in the field of fault diagnosis,while rough set theory is still in its infancy.In this thesis,we studied the related uncertainty problems in covering-based rough sets and fuzzy covering-based rough sets,and established a theoretical system of data analysis and mining for proposing more intelligent fault diagnosis methods in chemical process.Three types of chemical fault diagnosis problems were mainly considered,which are the Tennessee Eastman(TE)process,steam turbine unit and polymerization.The main work and contributions of this thesis are listed as follows:Ⅰ.Aiming at the problem of fault diagnosis under the condition of incomplete information,a method of eliminating redundant information from incomplete information fault features was proposed by using covering-based rough sets.It improved the accuracy of traditional fault diagnosis,and was applied to the fault diagnosis of chemical steam turbine unit.From the view of matrices,this thesis studied the problems related to the maximal and minimal descriptions in covering-based rough sets,and compares the matrix method with the traditional method by some open data sets in the machine learning library.The experimental results shown that the proposed matrix method could save the calculation time.By the matrix method of the maximal description above,a matrix calculation approach for maximal consistent blocks in the incomplete information system was proposed,which solved the problem of the complicated calculation when the data dimension is too large.Then,the original incomplete decision table was transformed into the maximal consistent block maximum description decision table by maximal consistent blocks.Based on the new decision table,a method of attribute reduction based on discemibility matrices was proposed.Finally,based on the proposed attribute reduction method under maximal consistent blocks,a fault diagnosis method of "maximal consistent blocks+intelligent classifiers" was constructed,which provided a method to solve the fault diagnosis problem under the condition of incomplete information.A simulation experiment was carried out for the fault diagnosis of steam turbine unit under the incomplete information condition.The experimental results shown that if the intelligent classifier is SVM(or Random Forest and Decision Tree),the accuracy of the proposed "maximal consistent blocks+intelligent classifiers"fault diagnosis method was 87.5%,while the accuracy of the fault diagnosis only using the above intelligent classifier was 75%.Hence,the accuracy was increased by 12.5%at least.Ⅱ.Aiming at the problem of fault diagnosis with complete information,a method of eliminating redundant information from fault features with complete information was proposed by using fuzzy covering-based rough sets.It improved the accuracy of traditional fault diagnosis,and was applied to fault diagnosis in TE chemical process.In theory,some problems of fuzzy rough set models based on fuzzy β-coverings were studied by the view of theory.Firstly,as a supplement to the existing concepts of reducible element and reduction in fuzzy β-covering approximate spaces,the concepts of I-reducible element and I-reduction were proposed.On this basis,an equivalence characterization between fuzzy β-minimal descriptions and β-reductions,and an equivalence characterization between fuzzy β-maximal descriptions and β-kernels were studied.Then,we generalized these concepts of one fuzzy β-covering approximation space to two fuzzy β-covering approximation spaces,and obtained new concepts and corresponding properties.On the basis of all the above results,relationships among one fuzzy β-covering and seven fuzzyβ-coverings induced by it,and the corresponding lattice structures were studied.In the application,an attribute reduction method based on fuzzyβ-neighborhoods was proposed under the fuzzy covering-based rough set model.On this basis,an intelligent fault diagnosis method of "fuzzy covering-based rough sets+SVM" was established.Finally,the TE chemical process was taken as the background,aiming at the following four states:normal,step fault(caused by the step change of process variable fault),drift failure(slow drift caused by the fault of chemical reaction kinetics)and viscous valve fault.A fuzzy covering information system could be established.Through the proposed attribute reduction method,23 fault features were determined from 53 fault symptoms in the fuzzy covering information system.By the "fuzzy covering-based rough sets+SVM" method of fault diagnosis simulation experiments,the accuracy was 86.57%,and only use the SVM method to get the accuracy was 72.5%.Ⅲ.On the basis of the two parts above,the generalized fuzzy covering rough sets and their application in chemical process fault diagnosis were further studied.On the one hand,based on the existing concepts of intuitionistic fuzzyβ-covering approximate spaces and intuitionistic fuzzy β-neighborhoods,as well as the first type intuitionistic fuzzy β-covering rough set model,we mainly studied their properties,and gave some new concepts and the second type intuitionistic fuzzy β-covering rough set model.Inspired by intuitionistic fuzzyβ-covering rough sets,the concepts of single valued neutrosophic β-covering and single valued neutrosophic β-neighborhood were put forward,and single valued neutrosophic covering-based rough set models were established.In order to solve the problem of multi-attribute group decision making,the single valued neutrosophic β-coverings and the single valued neutrosophic covering-based rough set models were extended to multi-granularity,and three kinds of multi-granularity single valued neutrosophic covering-based rough set models were constructed.On the other hand,under the condition of fault information,the group decision making method under intuitionistic fuzzy covering-based rough sets and another group decision making method under single valued neutrosophic covering-based rough sets were proposed.For polymerization fault diagnosis problem,there are three fault types:polymerization motor malfunction,polymerization deceleration machine malfunction,polymerization machine axis fault sealing,the polymerization kettle component fault and normal operation of polymerization,as well as the fault characteristics are:the polymerization reducer vibration value,operating pressure and mixing machine of decelerate of rotation speed and temperature of intuitionistic fuzzy information system and single value neutrosophic information system.The above decision methods were applied to the fault diagnosis of polymerization,and the final decision result of the proposed method was basically the failure of polymerization kettle motor.This is consistent with the results of other existing decision making methods.Therefore,the proposed decision-making methods under generalized fuzzy covering-based rough set for the polymerization fault are effective.In summary,based on the background of chemical process,this thesis adopted the method of combining theoretical research and experimental verification to further study the related problems of covering-based rough sets,fuzzy covering-based rough sets and generalized fuzzy covering-based rough sets(covering reduction problems,attribute reduction problems,etc.).On the one hand,the fault information was considered incomplete and complete fault information.T he attribute reduction methods were proposed to solve the probl em of feature selection in fault diagnosis under covering-based rough set(which is used to solve the incomplete information of fault diagnosis)and fuzzy covering-based rough sets(used to solve the fault diagnosis of complete information).The attribute reduction methods were combined with intelligent classifiers to improve the precision of the fault diagnosis.On the other hand,by using the generalized fuzzy covering-based rough set models,the multi-attribute group decision-making methods were established and applied to the fault diagnosis of chemical process.All of th ese provided theoretical and technical reference for intelligent fault diagnosis in chemical process. |
