Research Of Pattern Classification Method With Reject Option Of Process Fault Based On Hybrid Algorithm

Process faults often impaired the quality of outputs in the industrial system, sothe exact fault diagnosis revealed very important. However, the industrial systemflung down a big challenge to the fault diagnosis by its characteristics of complex andunobtainable process model. Meanwhile, the problem of faults overlapping each othercaused by small shift of process variables tended to make classification accuracy lowin the process of “discriminating” faults and “rejecting” new those. And the traditionalkernel-pattern classification methods had lost their strength for those faults data withnon-normal distributions, non-linear separability and overlapping each other. To beworse, when faults class labels were not available, the effect of “discrimination” ofpattern classification methods tended to be unacceptable without support from classlabels in a stronger form.Aim at the problems mentioned above, on the basis of summarization of solvingstate space problems, this paper proposes a series of new methods by usingmathematical analysis and constructing models. For the fault data following normaldistribution, new algorithm RS can effectively improve the performance of SVDDabout rejecting overlapping faults. The simulation results show the biggestimprovement can reach0.4260from0.5519to0.9779. new algorithm PFDA canlargely promote the ability of FDA to discriminate the overlapping faults. Thesimulation results show the biggest improvement can reach46.83%from52.65%to99.48%, and Relief F-KAKA algorithm also can largely enhance the KSDA from52.65%to93%totally40.35%.Without the assumption that normal distribution and linear separability and in thecase of labels available, this dissertation integrates Amari’s kernel function generationmethod and Kernel Alignment to form the KAKA algorithm, which can evade thesingularity problem from matrix inversion of Fisher criterion. Through the testing ofartificial data sets, KAKA promotes the recognizing rate from61.40%to99%at most,totally37.60%and superiors over KSDA. In the other hand, for labels inavailable,this dissertation proposes the new algorithm called SKK based on kernel functionoptimization using side information. Compared with traditional combination ofKFCM-F and KSDA, this algorithm gets better performance on promoting therecognizing rate from65.40%to94.50%at most, totally29.10%.The solutions to these problems not only break restraints on use of patternclassifictation methods of traditional assumption, but create new methods andtechnologies which have important theoretical value to extend process faults diagnosismethods.