Application of neural networks to fault diagnosis of multivariate control charts

Multivariate control charts are considered for the simultaneous monitoring of the mean vector and the covariance matrix when the joint distribution of process variables is multivariate normal The conventional multivariate quality control approaches evaluate the processes' control states based upon an overall statistic, such as Hotelling's T-2. As a result, the control chart can only give a total shift in controlled vector and can not point out directly whether the fault is arose from variation of subset or all of the variables. The application of traditional multivariate control chart is discounted for its fewer capabilities to guide the process adjustment. With the increasing of manufacturing processes' complexity and product quality requirement, multivariate quality control becomes necessity. Several modern multivariate control charts are proposed, such as modified multivariate Shewart (MMS) charts, mullivariate cumulative sum (MCUSUM) and multivariate exponential weighted moving average (MEWMA) charts etc. Each has some advantage as well as disadvantages. In this paper, by considering the cause-selecting problem as a pattern classification problem, a multilayer artificial neural network based model is proposed, which can diagnose fault patterns of process out-of-con trol state. Using with traditional multivariate control chart together, the model receives the process data as input when T-2 multivariate control chart gives aberrant signal, and produces fault pattern as output. The performance of the model is compared with MMS chart by numeric examples through considering possible variation combination. The results show that the proposed model has better performance especially when the number of quality variables or the number of out-of-control variables increases.