Impact of kurtosis on performance of mixture control chart patterns recognition using Independent Component Analysis and neural networks

Quality Management (QM) is one of the key supply chain processes. Previous studies have shown that QM provides a major contribution to the supply chain performance. One area that is essential in the managing of quality is Statistical Process Control (SPC). Control chart is one of the quality tools used in SPC for achieving process stability, controlling process and improving process by the reduction of process variations. Numerous techniques have been used to identify the presence of unnatural control chart patterns (CCPs); however, these approaches mostly focus on recognizing basic CCPs from a single type of unnatural assignable cause. In the situation where a mixture of CCPs exist, where more than one type of unnatural variation exist at the same time within the manufacturing process, this mixture of CCPs might be incorrectly classified. The Independent Component Analysis (ICA) technique has been applied recently to separate independent components from a mixture of two basic CCPs. This technique tries to maximize the statistical independence of the mixing components and uses kurtosis or a fourth-cumulant as a measure of non-Gaussianity, which implies statistical independence. Therefore, the CCP mixture separation performance and accuracy of using ICA greatly depend on kurtosis values and distributions of the basic CCPs. This paper will investigate the impact of the kurtosis values and distributions of basic CCPs on the effectiveness of the ICA-based approach in terms of being able to separate basic CCPs from a mixture, when using neural network as the pattern classifier.