Fault Diagnosis Of Industrial Processes Based On One-Class Support Vector Method

With the development of science and technology,industrial systems are developing in the direction of intelligence and intensification,and fault diagnosis methods which based on data have developed rapidly.Due to the high detection rate of single-class support vector method in industrial process fault diagnosis and its advantages for small samples and non-linear processes,support vector method has attracted widespread attention.At present,there are still some problems in the study of the method of single-class support vector machines(SVM):(1)due to the change of system operation,the problem of weak generalization of model hyperplanes in the application of SVM and imbalance of industrial data.(2)how to choose the model parameters in industrial processes fault detection based on support vector data description is an open question;(3)Traditional support vector data description method is to find the support vector based on the global information of the sample,local information that exist in the industrial data is ignored.(4)Most of the methods based on support vector data description have only completed industrial process fault detection,but how to further identify and isolate faults on this basis is also a problem to be studied.This paper studies the above problems,and the main work includes:An improved one-class support vector machine method is proposed and applied in industrial systems.The kernel density estimation method is used to improve the threshold selection of one-class support vector machine.Solving the weak generalization of the hyperplane of traditional model and imbalance of industrial data.The simulation experiment is carried out in three-capacity water tank system.The results show that the method reduces the false alarm rate while ensuring the fault detection rate under the imbalanced data distribution.Aiming at the problem of model parameter selection of the traditional one-class support vector machine method,the cuckoo search intelligent algorithm is introduced,and the objective function is defined by two important indicators in comprehensive fault diagnosis.Experimental verification is performed on the intelligent process control test platform.The cuckoo search algorithm is compared with the particle swarm search algorithm,and the results show the superiority of the proposed fault detection method.The method of Locality Preserving Projection Support Vector Data Description is proposed.Locality Preserving Projection is introduced in support vector data description method.The proposed method considers both the global structure and local structure of the data,the proposed method is verified in a numerical simulation system and Tennessee Eastman platform.Support Vector Center Deviation is proposed to solve the fault identification problems based on Support Vector Machines.The proposed method is applied in nonlinear industrial process to find the fault variables.Two cases,a numerical simulation system and Tennessee Eastman platform prove the effectiveness of the proposed method.