Abstract: The modern manufacturing industries have been accumulated volume and complexity quality data, data mining tools can be very beneficial for discovering interest knowledge and useful pattern in quality analysis. The traditional data mining tool is inefficient since following four aspects: the imbalance distribution, “curse of dimension”, mixed-type and data coupling. An attribution reduction algorithm based on equivalence relation is proposed for mixed-data feature selection and data preparation. Then an optimized kernel based hybrid manifold learning and support vector machines algorithm(KML-SVM) is presented for the imbalance distribution and “curse of dimension”. The manifold learning is used as an effective method to solve “curse of dimension” and the support vector machines is used to prediction and classification. Simulation based on the real quality data of a manufacturing process. The analysis of the results have shown the causality of the quality factors and given some suggestions on quality improvements. The simulation results have shown that the algorithm is effective and the ISOMAP kernel is an optimized kernel space.

Key words: Equivalence relation, Manifold learning, Quality improvement, Support vector machines