Research On Defect Recognition Of Flatness Shape Based On RBF Network And Support Vector Machine

Steel is the material basis for national economic development essential, as the main strip steel products, has become the raw electrical, automotive, civil and other industries widely used. Cold-rolled strip steel in a continuous production process, there will be a variety of plate-shaped defects affect the subsequent rolling and strip quality. So, how to shape detection data and pattern recognition methods to get strip-shaped defect types and characteristics of parameters to improve shape recognition speed and accuracy, so that subsequent rolling out of the plate-shaped strip flatness to meet user requirements is a very critical issue.This thesis focuses on the plate-shaped defect pattern recognition research, the model identified by plate-shaped characteristic parameters, analyze the advantages and disadvantages flatness defect identification methods, and a new method proposed flatness defects identified on this basis.(1) Because the traditional method of least squares does not recognize complex plate-shaped, while the weights of BP neural network board shape recognition method is difficult to determine. The genetic algorithm optimized BP neural network used in the plate-shaped recognition. Through simulation analysis, genetic algorithm optimization recognition model, the recognition accuracy is higher than the BP network, but recognition was slower, does not apply to online flatness recognition.(2) The K- means clustering RBF neural network applied to the plate-shaped defect recognition, to determine the center of radial basis function C, variance σ by K- means clustering algorithm. As the network input more will fit complexity increases, the difference between the weighted Euclidean distance will reduce the amount of input from 20 to three. After simulation, the GA-BP, GA-RBF and K-RBF identification methods were compared, the results show that K- means clustering RBF method can accurately identify the six common types of plate-shaped defects, high recognition accuracy and speed fastest.(3) Because of least squares support vector machine(LSSVM) under nonlinear, small samples of pattern recognition is better, will be used in the plate-shaped defect LSSVM recognition. The simulation showed that, LSSVM can accurately identify the plate-shaped characteristic parameter identification error is slightly lower than the five groups of samples K-RBF method for cold-rolled sheet-shaped line adjustment.Through these studies, K- means clustering RBF method and least squares support vector machine method for identifying defects flatness better than now commonly used BP neural network, in line with the plate-shaped line recognition results are accurate, fast requirements, can be applied to practical engineering.