A Study On Autocorrelated Process Control

The assumption of traditional statistical process control (SPC) techniques is thatprocess data are independent and have the same distribution. So SPC does not fit toautocorrelated process. The use of traditional Shewhart control chart on autocorrelatedprocess will show a large amount of false alarms which will lead to wrong decisions,and it will also bring some misleads on quality management and quality control. Allthis makes the effect of control charts and the control level on autocorrelated processfall. Techniques which are used to identify shifts in correlated parameters are neededto serve the same functions as SPC control charts.Neural networks are a potential tool for identifying shifts in correlated processparameters, as data independence is not an assumption of it. In this research, timeseries based residual control charts are used on control of process parameter valuesfrom AR(1) time series models with varying values of the autocorrelationcoefficient φ and different shifts. Then BP neural networks and SOM neuralnetworks are used in the same simulated process, and a contrast among differentmethods is provided. At last, all the methods are applied in a case of a real chemicalprocess.As the results reveal, residual control charts are easy to use in manufacturingprocess, but there are many limitations. On many occasions, residual control chartscan't recognize process shifts correctly;BP neural networks were successful atseparating data that were shifted one, two and three standard deviations fromnon-shifted data for generated process data, but its calculation theory makes it notapplicable in real manufacturing process;SOM neural networks' ability inrecognizing shifts of autocorrelated process is comparatively low, but it is applicablein real production process,and the case in the paper shows that SOM does good in realautocorrelated process controlling. In other words SOM is a potential good method inautocorrelated process parameters control.