A Support Vector Machine Based Multi-kernel Method for Change Point Estimation on Control Chart

Despite the abnormal patterns recognition and mean shift size estimation of control chart signals could provide some evidence for statistical process diagnostics, it do not reveal the real time of the process changes, which is essential for identifying assignable causes and ultimately ensure stability of process. In this paper, a support vector machine based multi-kernel (MK-SVM) method for change point estimation on control chart is proposed. For this purpose, the moving window analysis is introduced to decompose the whole process sequence features into multiple time sub-sequences, and different types of kernel functions are combined together by using kernel method, which is mapped into a new feature space to form the multi-kernel function of SVM. Then each characteristic of the sub-sequences is regarded as a determined pattern to be recognized through the proposed model. We use the cross-validation method to search the optimized parameters of MK-SVM. Multiple sets of experiments are used to verify this method. Finally, a case study about the coating process of production lines is conducted to evaluate the performance of the proposed approach, results reveal that the proposed scheme is able to effectively estimating the time of change-point and outperform the commonly used approaches.

1. Introduction

Control chart is one of the most effective tools in statistical process control (SPC), which is the most widely used to reveal abnormal variations of monitored measurements, as well as to locate their variation width. However, it could not indicate the real time of the process changes. This information is of great importance for the quality engineers to search for the assignable causes, it reveals the potential quality problems as well as to take some necessary corrections and adjustments to bring out-of control process back to the in-control state [1]. Moreover, due to many factors thus as man, machine, material, method, environment, measured (5MIE) existed in the manufacturing process, it leads a great difficulty to detect the real time of the process changes. Thus it is of great value to explore the change-point detection method by using machine learning theory.