Applied Research On Quality Diagnosis By Combination Classifier Based On The Gradient Descent

In modern time, that treated quality as life, quality has been the means of winning the market and making profits. It is in this context, when the statistical process control was produced. As a means of production process monitoring and controlling, quality diagnosis has become one of current hot research topics. In so many tools which can monitor the validity of the process, Control chart is most classical and effective. In early practice, as a result of the limitation of technology, people can only monitored a few quality characteristic which obviously influence the quality of product. During this time, because the theory assumed that all of the variables were independent of each other, so people could monitor each variable respectively by using the traditional control chart. However, at this time, the diagnostic methods could not really achieve the aim of improving the quality of the product. With the development of the practice, practitioners began to use the associated control chart to monitor production. Because there were so many combinations between the quality characteristics, it was difficult for people to implement in practice.In the academic field, when the Hotelling T2control chart give signal that sample is out of control in multivariate production process, it has become an urgent problem needed to resolve that how to determine the specific variable out of control. Domestic and foreign scholars have put forward many quality diagnosis methods, such as diagnosis method based on principal component analysis, MYT decomposition method, diagnosis method based on neural network and diagnosis method based on support vector machine and so on. Although there are many scholars have proposed many multivariate quality diagnosis method, there are always some unsatisfactory places.Based on the extensive summarizing the predecessors’ research results, we select the classification and regression tree with two layer tree deep and four leaf nodes as the base functions, then combine the steepest descent method and the combination classifier. Through the continuous optimization the parameters of the next generation base classifier, we can get many classifiers which can better complementary. We proposed the combination classifier which based on gradient descent method, and gives the diagnosis method. Through the case study and comparison, we find our method excel the multilayer perception(MLP) and the multivariate Shewhart chart(MSCH) in the diagnosis accuracy. Finally, we make some improvement on our method, and get some better effect. From the perspective of practice, this paper research has the significance of directing production practice. Besides that, all the example used in this article are from the previous literature, we also make a comprehensive comparison with other research results, this paper also can enrich the theory of quality diagnosis.