A Study Of Support Vector Machines Based Fault Diagnosis And Its Applications

With the growing increase of modern industrial system complexity, the data-driven based faultdiagnostic approach has become one of the hottest subjects among up-to-date researches. Beingdifferent from classical approaches which require precise mathematical models, the data-drivenapproach approximates the function between system’s input and status by constructing a ‘balck-box’model using the process data. And this makes the data-driven approach more suitable for diagnosingthe complicated nonlinear systems. The method using Support Vector Machine is an important kind ofdata-driven approach. It overcomes the problem of dimensional disaster and avoids the local minima.Moreover, as it employs the principle of Structure Risk Minimization into the process of classificationhyperplan construction, the SVM-based method requires smaller amounts of training samples thanconventional methods. Hence, it is with high value for both theoretic studies and real applications. Inthis thesis, three common issues have been investigated for SVM-based fault diagnosis, including:(1) How to extend the binary SVM classifiers into a multiclassifier, and employ the multiclassiferinto the fault isolation: decision Directed Acyclic Graph SVM (DAG-SVM) is one of the mostefficient extending strategies for SVM. However, it has a division bias while deciding. Focusing onthis problem, a node-refined approach has been proposed to optimize the DAG-SVM decidingstructure. It obtains a structure with the smallest probability of misclassification risks. Further,considering the differences between losses that brought by different misclassifications, the totalmisclassification loss is treated as the object of optimization in my thesis. A misclassification lossminimization SVM has also been proposed. This approach attempts to find the structure with the leastmisclassification loss, but not the least misclassification probability. Consequently, it is morereasonable for real applications. Both of the two proposed approaches were applied to the diagnosis oftransformer faults in this thesis, and satisfactory results were obtained.(2) How to reduce the deciding time consumption and accelerate the response speed while SVMapproach is implemented for online fault diagnosis: a label-based online diagnosis system has beenproposed in this thesis. It divides the system into two parts: the ‘offline learning’ part, and the ‘onlinedeciding’ part. By a deep mining of label information, the ‘offline learning’ part finds the optimalfeature subspace. Then the ‘online deciding’ is made within this subspace. As the number of featuresinvolved in the subspace is much lower than it is in the original data samples, the dimension of testingsample in the ‘online deciding’ part has been reduced, and hence the computation has been reduced. These reductions lead to the acceleration of deciding speed.（3）How to repress the migration of classification hyperplan,which is brought by the imbalancebetween positive and negative traning samples: A self-tuning SVM has been proposed in this thesis.This approach estimates the quantity of information that each sample contains using the AdaBoostmethod. And it generates a penalty factor for each sample according to the estimation. Meanwhile, thetotal misclassification losses of both positive samples and negative samples are also guaranteed to bethe same in self-tuning SVM. Hence, the proposed approach has efficiently repressed the migration ofclassification hyperplan.After the discussions of several advanced SVM approaches for fault diagnosis under normalconditions, the thesis also attempts to improve and optimize the performances of SVM approachesunder some special conditions, including:(4) How to improve the diagnostic accuracy in condition that one or more types of faultysamples are missing: The SVM-based fault diagnostic approach is sensitive to the samplecompleteness. There will be a big migration of fault region divisions if one or more types of faultysamples are missing when the SVM method is implemented to fault diagnosis. Hence, the supportvector data description (SVDD) was introduced in this thesis in order to refine these fault regiondivisions. SVDD has successfully reduced the sensitivity of sample completeness when classifiers areimplemented for fault diagnosis, and efficiently improved the diagnostic accuracy.(5) How to make effective fault detections in condition that only normal samples were obtainedand there were no faulty samples: A flexible Support Vector Regression (F-SVR) approach has beenproposed to approximate the function between system inputs and outputs in this thesis. By theanalysis of residuals, the fault detection could be made. Compared with conventional support vectorregressions (SVRs), the F-SVR is capable to make an automatic setting of the regression parameters.It yields a good generalization ability while maintains a good learning ability. The proposed approachhas also been applied to the fault detection of high frequency power supplies, and satisfactory resultshave been obtained.