Classification techniques for control chart pattern recognition: A case of metal frame for actuator production

• Main Authors: Wimalin Laosiritaworn, Tunchanit Bunjongjit
• Format: Journal
• Published: 2018
• Subjects: Biochemistry, Genetics and Molecular Biology; Chemistry; Materials Science; Mathematics; Physics and Astronomy
• Online Access: https://www.scopus.com/inward/record.uri?partnerID=HzOxMe3b&scp=84888105458&origin=inward; http://cmuir.cmu.ac.th/jspui/handle/6653943832/52184
• Institution: Chiang Mai University

Statistical process control (SPC) plays a significant role in hard-disk drive manufacturing as there is a crucial need to constantly improve of productivity. Control chart is one of the SPC tools that have been widely implemented to identify whether nonrandom pattern caused by assignable cause exists in the production process. Decision rules are usually used for detecting nonrandom patterns on control chart. However, recent research has shown that these rules had tendency of producing false alarm. This is a problem occurred in the case study company, who is a manufacturer of metal frame for actuator. The company is adopting technologically advanced equipment for its quality assurance system and computer software for data analysis and control chart. Currently, the company use decision rules for detecting nonrandom patterns on control chart - for example, if 6 or more consecutive data inputs found to be in an increasing or decreasing order, these data contain trend pattern. In attempt to improve the accuracy of data analysis, this research investigated the application of 3 classification techniques, namely neural network, k-nearest neighbor and rule induction, in discretion of nonrandom patterns. By considering the control charts of 3 different product lines, 3 types of nonrandom patterns, which are Trend, Cycle and Shift, are to be observed. Based on the real data inputs, the percentage of accuracy in error detection by each technique of each product line is compared. It is found the accuracy of k-nearest neighbor is highest with the percentage of correctly prediction between 96.99 - 98.7%.