Based On Support Vector Machine Ferrography Image Recognition Technology

With the safe and reliable operation of large-scale petrochemical machineryand equipment has become more complex and important, attracted the attentionof many people. Ferrographic analysis as one of the effective means of oilmonitoring, in addition to monitoring the physical and chemical indicators oflubricants, mainly through the wear debris morphology analysis to determine thedegree of friction and wear of mechanical equipment and failure type. Fortraditional wear particles morphology analysis were observed, judged and madea diagnosis report by experienced staff, which not only increases the burden onthe staff, but also with the human factors. For this reason, along with thecontinued development and application of computer technology, researchers willuse computer technology to analyze ferrographic image by this way can promotethe rapid development of ferrographic analysis techniques．With the extensive application of neural network technology for smallsample classification and regression support vector machine (SVM) in machinelearning has also exhibited more advantages, and are widely used in imagesclassification and recognition, voice recognition, machinery fault diagnosis andso on． There are five kinds typical wear particles about150ferrography imagesincluding severe sliding, spherical, cutting, fatigue were chosen to analyze. Inorder to achieve ferrographic images classification, feature information of weardebris images should be extracted firstly, because the feature extracted quality ofimages plays a vital role on the final recognition results. The wear particles canbe better segmented from the image by K-means clustering segmentation，extraction region growing, mathematicl morphology erosion and expansionprocess. Using matlab software to extract wear particles morphology parameters:shape, size, texture and color, about17parameters and150groups of data. Inorder to obtain a better classification accuracy, SVM parameters should beoptimize to get optimal parameters. Through analysis of the domestic andforeign issues in depth, genetic algorithms was used to optimize SVMparameters that were penalty factorC and kernel parameter g for getting the bestfitness and multi-class classifier including optimal parameters and using theclassifier to predict the sample data. Finally, the recognition accuracy of testedwear particles sample can reach more than90%. The result showed that basingon genetic algorithm optimization of SVM ferrography images recognitionmethod in this paper achieved good effect．...