Research On The Methods Of Fault Diagnosis In Process Industrial Based On New SVM Algorithms

Fault diagnosis in process industrial is of great significance to reduce accident and economic loss. However, some existing method effect is not very ideal applied in faults diagnosis of the process industry process. Because methods of fault diagnosis based on machine learning and artificial intelligent do not depend on precise mathematic model, experts and scholars pay great attention to them in the process industry fault diagnosis field. Support vector machine (SVM) showed excellent classification performance based on the statistical learning theory and it becomes a one of the research hotspots of in the field of fault diagnosis.This paper is focus on support vector machine (SVM) applied in the process industry of fault diagnosis and it is blast furnace of fault diagnosis and chemical Tennessee Process (Tennessee Eastman Process, TEP) fault diagnosis as background, and it puts forward some new algorithm of the support vector machine (SVM) what are innovation exploring to process industry of the fault diagnosis.This paper studies on selection and optimization parameter of support vector machine (SVM) problems and put forward the new method of fault diagnosis based on the Nearest Discrete Particle Swarm Support vector Machine (Nearest Neighbor-Discrete Particle Swarm Optimization-Support Vetor Machine, NN-DPSO-SVM). The method for the training set with noise removed some sample with small effect by nearest neighbor rule and making the training set cut. At the same time, feature selection and support vector machine (SVM) parameters are optimized by the improved discrete particle swarm optimization algorithm and what improve the support vector machine performance and identification accuracy. In the evaluation standard of support vector machine performance based on the study of the problems in depth, this paper presents a new method about fault diagnosis that is cost-conscious least square support vector machine. This method concludes a new cost-conscious formula and makes it as fitness function. The fitness function can fully take into account precision and speed of support vector machine and the number of support vector and measure better the generalization ability of least square support vector machine and optimize the feature vector and related parameters of the least square support vector machine (S VM) and shorten the classification time.This paper studies on multiclass classification method of support vector machine (SVM) and put forward a new gradual change binary tree multiclass classification method. This method improves one-against-all multiclass classification of SVM and overcome its shortcomings that classification number is not unbalance during classification. The method is based on the distance formula among the fault samples and builds the SVM binary tree based on the characteristics of the fault samples and it is accord with the actual requirement of the fault classification. It is applied to the fault diagnosis of blast furnace experiment and there has a very good classification effect.Different states data sample of the process industry process is often not balanced and the number of normal state of fault samples are usually more than the number of fault sample. This kind of imbalance data classification is separated by standard support vector machine (SVM) and obtaining a classify plane is not optimal and will make classifier generalization performance decline. In order to solve this problem, this paper studies deeply the unbalanced data classification method and puts forward two new methods. First, this paper puts forward a new adapting support vector machine (SVM) algorithm. The algorithm consists of filling less-class algorithm and complementary algorithm and they mainly use adapting transductive support vector machines (adapting transductive support vector machine, ATSVM) and edited nearest neighbor principle. The filling less-class algorithm selects the meaningful test data of less-class. The samples can be used to supplement the deficiency of the training samples. However, the data also includes noise data. Therefore, edited nearest neighbor principle remove noise samples. The complementary algorithm selects the valuable less-class samples and more-class samples and adds them to the training set. The methods removing noise data is similar to the filling less-class algorithm. Both of two algorithms add test samples so that they keep the balance between more-class samples and less-class samples. Secondly, this paper put forward optional support vector machine (SVM) new algorithm. This method is through clipping training set and joins unlabelled data to the training set and supplements the lack of training samples to make the training set more representative and closes the sample sizes gap between the less class and the more class. Fixed binary tree of SVM classification method is based on the construction of the characteristics of the interdependence among faults.It not only shortens the training time, but also improves the classification ability.To realize the early fault prediction, the paper puts forward a new weighted support vector regression machine algorithm and what can predicte abnormal condition through the parameters change.A large number of simulation experiment based on these new algorithms what are put forward in this paper were compared with other SVM fault diagnosis methods. The experimental results show that these new algorithms have higher this classification accuracy and faster the classification speed and better generalization ability.