216 Artificial Neural Network in Applying Multi Attribute Control Chart for AR Processes Seyed Taghi Akhavan Niaki, Ph.D., Professor, Shirin Akbari Nasaji, BSc. Department of Industrial Engineering Sharif University of Technology (SUT) Tehran, Iran 1458889694 e-mail: niaki@sharif.edu , shirin.akbari@gmail.com Abstract— Quality characteristics are subject of both manufacturing and service industries, which include not only the variables but the attributes as well. In Quality Control area substantial research has been done for Auto-correlated variables; however, no attempt was done for Auto-correlated attributes. Ignoring the autocorrelation structure in constructing control charts cause the in-control run length to decrease, and the false alarms to increase as such. In this article we develop a new methodology based upon the modified Elman neural network capabilities to overcome this problem. Moreover, instead of back propagation, simulated annealing is suggested as an alternative training technique that is able to search globally and in order to generate random AR vector we develop another artificial neural network based on ARTA algorithm. We present a simulation experiments and compare the performance of the proposed methodology with the other control methods of multi-attribute processes. The result of the simulation study is encouraging. Keywords-component; Multi attribute control charts, neural network, Autoregressive, ARTA I. INTRODUCTION Statistical control of many continuous processes deals with data that exhibit some degree of autocorrelation. In a control charting method development, the presence of autocorrelation results in violating one of the key assumptions, namely serial-sample independency. Violating the independency assumption affects both the in and out-of- control average run length (ARL) of the control charts and makes them unreliable [3]. In general, there are two broad categories in statistical control charting methods; variable and attribute control charts [12]. The problem of monitoring auto-correlated processes in variable control charts motivated many researchers. Alwan [1] studied the effects of autocorrelation on the performances of the Schewhart control charts and found that there is an increased probability of both false and genuine alarms when correlation is present. Kalgonda and Kulkarni [9] proposed a new multivariate chart to detect and diagnose mean shifts in presence of autocorrelation for observations modeled by vector autoregressive of order one (VAR(l)) process. Many researchers have proposed several approaches to take into account autocorrelation in the control charts. One of the main approaches to deal with autocorrelation, which was introduced for the first time by Alwan and Roberts [2], 978-1-4244-5586-7/10/$26.00 ©2010 IEEE is to use a residuals control chart. Residuals are the difference between the real and the forecasted values of the mean vector of the process variables. They proved that when they used an appropriate model to forecast the variables, calculated residuals would be serially independent, so it was possible to apply the common control charts for the residuals. Despite the fact that the multi-attribute monitoring has many applications, almost all researchers have focused on the first category of control charting and only a few methods have been proposed to monitor multi-attribute processes. Furthermore, in many instances where we may not need exact measurements it is easy to collect correlated discrete- type data. Patel [17] proposed a Hotelling-type χ2 chart to monitor observations from multivariate Binomial or multivariate Poisson distribution. In another research, Lu et al. [11] addressed the statistical design of multi-attribute control charts, where they proposed a multivariate np-chart (Mnp chart) to develop Shewhart charts based on an X statistic. Furthermore, they showed that this statistic reduced type II errors better than individual np charts since the correlation of attributes was taken into account. Jolayemi [8] developed a model for an optimal design of multi-attribute control charts for processes with multiple assignable causes. In a more recent research in this area, Niaki and Abbasi [13] based on a simple approach that almost eliminates the existing correlations between the attributes, presented a methodology to monitor multi-attribute processes, and presented a rectangular region for the monitoring purposes. Witnessing the increasing capability of artificial neural networks (ANN) in modeling real systems, there is a great interest in ANN for quality monitoring. Accordingly, Larpkiattaworn [10] proposed a back propagation neural network (BPNN) for bi-attribute Binomial process control chart using the assumption of a positive correlation providing a large enough sample size. In another research, Niaki and Abbasi [14] employed a neural network to detect and classify mean shifts in multi-attribute processes. They applied the network to monitor not only the proportions of several types of defects, but also the number of different defects in a product online. Niaki and Davoodi [15] considered a production process in which there were several auto-correlated stages and in each stage, there were several correlated quality variables to be monitored. In order to control the resulting auto-correlated multivariate-multistage process, they designed a single fully connected feed-forward four layers ANN. To the authors' best knowledge, a standard method for constructing a multi-attribute control chart for auto- Volume 5 Authorized licensed use limited to: Sharif University of Technology. Downloaded on June 06,2010 at 03:34:08 UTC from IEEE Xplore. Restrictions apply.