Research On Abnormity Diagnosis In Multivariate Statistical Process Based On Ensemble Learning

Now the century is the century of quality. High quality products and services are the consistent pursuit of countries and enterprises. The improvement of product quality can’t do without the control of the production process. With the continuous improvement of product complexity, it is often necessary to monitor two or more related quality characteristics simultaneously in production process so that we can detect the abnormality and then take some corresponding measures to correct the process. At last, we can achieve the purpose of improving the quality of products. However, the control charts which are designed for monitoring the production process can detect an unusual event but do not determine which variable or group of the variables has caused the out-of-control signal. In order to resolve this problem, this paper proposed a method of ensemble classifier model based on Bagging and classification and regression tree, which is used to identify the control signals of mean shifts given by HotellingT2 control chart. Three simulation experiments that include two, three and four variables respectively are adopted. We selected suitable training times, which will exert an influence on the result of classification, through lots of experiments. The validity of this method is proved by the results of simulation experiments and comparison with other methods. At the same time, the number of samples in training set and the correlation between variables that affect the efficiency of the method were analyzed. Then, we further use this method to monitor and identify the mean shifts in multi variate statistical process on-line by the application of the moving window method. The efficiency of abnormity identification is measured by the average run length, and the accuracy is measured by the correct classification rate. In the simulation experiment, an appropriate window size is selected through a lot of experiments. Experimental results verify the effectiveness of this method. For the average run length, comparison with other traditional control chart methods show that the method can quickly identify abnormity and have high efficiency of identificatioa The results of this study have important theoretical and practical significance for the multivariate quality control.