Two-Stages Monitoring and Diagnosis of Mean Shifts in Bivariate Correlated Process using an Integrated MEWMA-ANN Scheme

Production Automation and Control

by yeimzai on November 3, 2010 - 15:59

Two-Stages Monitoring and Diagnosis of Mean Shifts in Bivariate Correlated Process using an Integrated MEWMA-ANN Scheme

Ibrahim Masood*, Univ Tun Hussein Onn Malaysia; Adnan Hassan, Universiti Teknologi Malaysia

Abstract: Various artificial neural networks (ANN)-based pattern recognition schemes have been developed for monitoring and diagnosis of mean shift variations in bivariate processes. These schemes generally perform better in classifying mean shifts and provide more diagnosis information compared to the traditional multivariate statistical process control (MSPC) charts. However, some disadvantages in terms of reference multivariate/bivariate patterns and excess false alarms may restrict attention on the scopes and development in this area. Therefore, this paper aims to investigate two-stages monitoring and diagnosis for some reference bivariate correlated patterns using an integrated multivariate exponentially weighted moving average (MEWMA)-ANN. Feature-based input representation was applied into an ANN training towards improving discrimination capability between bivariate normal and bivariate mean shift patterns. Besides comparable diagnosis performance, the proposed scheme has resulted in better monitoring performance with smaller false alarms and faster detection compared to the raw data-based ANN scheme.

Download paper

Download presentation

2-stage method

November 22, 2010 - 16:54 - by fahmy Dear Authors, Thank you for your participation in the conference. I got one question. What is the significance of using the two stage and applying the NN, while both the NN and the MEWMA are both having thier input from the system measurments?. Regards A. Fahmy

2-stage method

December 26, 2010 - 03:05 - by yeimzai Thank you Fahmi, I am sorry for very late reply. Two problems in the traditional multivariate SPC charting schemes (T2, MCUSUM, MEWMA) are: (i) unable to identify the source variable(s) of mean shifts, and (ii)produce relatively high false alarms compared to the univariate SPC charting (Shewhart, CUSUM, EWMA). The existing pattern recognition schemes (Zorriassatine et al, 2003;Niaki and Abbasi, 2005; Guh, 2007; Yu and Xi, 2008; etc.) applied the ANN-based models either for:(i)only identifying the source variable(s) from the MSPC out-of-control signal, or (ii) both monitoring and diagnosing bivariate process mean shift through direct continuous recognition. However, the existing approach does not reduce false alarms (ARL0 still maintain = 200). Therefore, using two stage monitoring approach, the ANN-based model, namely, feature-based ANN was applied to reduce false alarms (ARL0 increased to 350+). The other advantages are:(i) Faster detection capability for bivariate process mean shifts with shorter ARL1 results, and (ii)Correctly diagnose the source variable(s) of mean shifts with high recognition accuracy results. The MEWMA control chart was selected for this study rather than T2 control chart because it is sensitive to small mean shifts (shorter ARL1).