(429e) A Deep Learning Vision System for Classification of Manufacturing Defects | AIChE

(429e) A Deep Learning Vision System for Classification of Manufacturing Defects


Hanselman, C. - Presenter, Eli Lilly and Company
Tiwari, A., Eli Lilly and Company
Venugopal, M., Eli Lilly and Company
Guan, Y., University of Michigan
Comparini, E., Eli Lilly
Zhang, B., Eli Lilly and Company
Prucka, W. R., Eli Lilly and Company
Manufacturing lines for the assembly of pharmaceutical autoinjectors are highly automated and rely on visual inspection stations to efficiently and safely process a large volume of syringes. As pre-filled syringes enter the inspection line, there are a variety of subtle defects in the flange (e.g., chips, cracks) as well as other processing flaws (e.g., missing components) that can make the assembly unfit for use or otherwise problematic for continued processing. Currently, there is a highly-engineered inspection algorithm that is manually tuned for identifying defects from camera images. The existing approach is able to robustly detect important defects, however, a relatively large percent of rejected products are in fact false positives (e.g., cosmetic defects, lighting irregularities). In this work, we have developed a convolutional neural network approach for classifying defects in syringes with comparable sensitivity and higher selectivity than the previous algorithm. We found that preprocessing and noise reduction were critical for training meaningful models from available images. Additionally, we highlight a series of infrastructure challenges that were overcome during image acquisition and algorithm deployment to transfer our work to and from a good manufacturing process (GMP) inspection line. We also illustrate the importance of algorithm-architecture co-design when developing machine learning models in order for the solution to be practically deployed on the available edge computing device (i.e., the vision processor). This work highlights the use of machine learning in conjunction with traditional computer vision to classify defects beyond what can be identified by expert intuition.