On-line adaptive Bayesian classification for in-line particle image monitoring in polymer film manufacturing

Contaminant particles suspended in polymer melts flowing through an extruder can result in film defects which ruin film performance and appearance. In-line monitoring of the polymer melt using a specialized camera system provides images which can be used to anticipate and potentially diagnose the cause of such defects. However, image interpretation is sensitive to changes in image quality. Development of a practical method for adapting to such changes during an extrusion operation and automatically distinguish images containing contaminant particles from those that do not, was the objective of this work. This was accomplished off-line by using a database of about 6000 in-line acquired images and a very recently developed adaptive machine learning method employing a Bayesian model. Performance, robustness, structure complexity and computational time considerations are examined.