USING POLYNOMIAL SMOOTHING AND DATA BOUNDING FOR THE DETECTION OF ADVERSE PROCESS CHANGES IN A CHEMICAL PROCESS

This paper focuses upon the problem of detecting outliers in a time se ries used to model a production process in the chemical industry. Sign ificant deviations from the underlying time series pattern, i.e. outli ers, indicate an adverse process change or out-of-control situation re lative to the model. The underlying process is modeled using either le ast squares moving polynomial fit smoothing based upon the Savitzky-Go lay algorithm21 or data bounding. This makes any outliers in the origi nal data more salient when compared to the smoothed graph. Thus outlie rs can be detected earlier while the process output is still within st andard control limits and product specifications. The proposed algorit hms improve upon and complement the conventional control chart, partic ularly with interdependent observations. The process control capabilit ies of these methods were successfully tested on an autocorrelated dat a set taken from a chemical production process with known adverse proc ess changes and assigned causes. These algorithms should be of assista nce to the chemical engineer or industrial chemist involved in process and quality control.