Search

In order to overcome the effects induced by the complexity of sample pattern, noises and the non-integrality of information, a modeling method called rough SVM is put forward by applying the theory of rough set to binary hyper-sphere SVM and employing the advantages of rough set and SVM. Rough set has the capability of describing uncertain, non-integrated data and complex pattern, but it has no good learning capability and can not ensure the generalization of classification model. SVM possesses the outstanding performance of generalization, but it has bad capability of modeling uncertain data. Classification results in this paper are divided into positive field, boundary field and negative field, by which the uncertain degree of classification results is judged. Through adjusting the parameters, the width of boundary and the ratio of outliers allowed to the model can be adjusted, and flexibility of the classification model can be improved. Simulation results show the availability of this modeling method. [ABSTRACT FROM AUTHOR]

Because the detect recognition characteristic extracted from the welding image bas been seriously interfered by noises, and the accuracy of existing recognition algorithms is low, a defect recognition algorithm integrating rough set (RS) with fuzzy neural network (FNN) is presented in this paper. Firstly, the fuzzy C-mean (FCM) clustering algorithm was adopted to discretize the attributes of samples, and RS was used to reduce the attributes of sample data and obtain decision rules, then π function was used to fuzzified the attributes of samples according to the center and radius of clustering to overcome the problem that RS is sensitive to noises. Then, the obtained reduced fuzzy decision rules and fuzzy logical inference were used to ascertain the structure of FNN, and dependent factors together with antecedent coverage factors were employed to determine the initial parameters of network. In consideration of the reliability of the data in the samples, the weighted cost function was used to adjust model parameters. The simulation result shows that this algorithm can solve the problems, such as the uncertainty of sample data caused by noise interference in the process of classification and the difficulty in determining the structure of FNN, and can greatly improve the recognition capability of welding image defects. [ABSTRACT FROM AUTHOR]