Study Of Support Vector Machine In Recognition Of SPC Control Chart Patterns

SPC control chart is one of important and commonly used tools for the modernquality control. According to the state of control chart, abnormal situation and thepotential quality problems in the production process can be judged and revealed.Generally, control chart patterns are divided into two modes, including normal patternand abnormal pattern. Different abnormal patterns represent different productionanomalies. So it’s necessary to recognize control chart patterns to find the reasonsbehind of different abnormal situation, with improving product quality and productionefficiency. In view of this, the recognition of control chart is of great importance.The traditional SPC recognition methods are based on large sample condition,and when the neural network methods of diagnosing process abnormalities are used, alarge number of training samples are also needed. These two methods can work wellwhen samples are large. But when samples are limited, the drawbacks of the twomethods are presented. Therefore, how to achieve quality control effectively under thecondition of small samples has become one of the main research problems in qualitycontrol and diagnostic. In recent years, support vector machines is becoming a newintelligence method of artificial intelligence research, which can deal withpattern classification better under the condition of small samples. In this paper, SVMare used, and methods for control chart patterns recognition under the condition ofsmall samples are studied. The main work can be summarized as follows:1. In the thesis, theory of SVM and MSVM were introduced and classifier designmethod of "1-v-1" was proposed. On the base of analyzing the impact of theparameters of SVM model on classification performance, application of PSOparameter optimization in SVM parameters was proposed, and PSO-SVM classifierwas built.2. Different identification features including the original features, principalcharacteristics, statistical characteristics were applied into recognition of control chartpatterns. Simulation experiments and analysis were also conducted.3. In order to improve the recognition accuracy, two improve methods wereproposed on the base of shortcomings presented in previous studies. Considering therelevance and redundancy between statistics features extracted, a method based onimprove sequence forward selection method was presented. In order to avoid missing of important information in control chart and identify control chart better, a methodbased on the fusion features were proposed, a kind of fusion features included theoriginal feature and PCA feature; another included original feature and statisticalfeature. In addition, comparison was conducted between different recognitionmethods. Experimental results show that the improved methods do improve therecognition performance of control charts effectively.