The traditional fault detection suffers from complicated process, low accurate ratio and off-line implement. The improved methods of defect recognition by artificial neural networks (ANN) can lead to the problems of overfit and bad generalization because of finite samples. With a view of data mining and technique parameters directly, the new approach using support vector machine classification algorithm after removing redundant parameters by rough set theory and eliminating noise of data to identify the defects is discussed. The results of a experiment show that unlike conventional and ANN recognition methods the new technique performs better than conventional evaluation ones with advantages of high efficiency, lower cost, easy implement on-line, excellent generalization and so on. The approach provides a novel technique means for nondestructive defect identification of various products.