Data mining methods for quantitative in-line image monitoring in polymer extrusion

This thesis is based upon the hypothesis that data mining methods, when combined with existing and new monitor interpretation methods, can enable in-line image monitoring of polymer film extrusion to be used for high throughput analysis. Data mining refers to a wide variety of methods available for deriving information from data, especially large quantities of data. The specialized camera system used to obtain the images was a "scanning particle monitor" that can obtain images across the direction of the flow of the polymer melt. The two objectives necessary to show the validity of the hypothesis were accomplished: automated quantitative image interpretation methods for detecting the presence, number, size and concentration of contaminant particles were devised and experimentally verified using thousands of images; these methods were combined and implemented to enable the interpretation to adapt to dynamic environments. A new, very advantageous method of image thresholding was derived (termed "maximizing the minimum particle size"). Bayesian classification was found to be superior to other classification methods when both error rate and false negative rate was considered for detecting images containing contaminant particles. Decision tree was found to be superior to other classification methods for distinguishing in-focus from blurry images of particles. The "intelligent learning machine" method, based upon supervised learning to achieve adaptability, was used with Bayesian classification and the new thresholding method to provide the needed adaptive, real-time interpretation algorithm.