AUTOMATED QUALITY CONTROL IN MANUFACTURING
The No. 2 automotive headlight company in the world approached us with the challenge of automating quality control in order to free human resources for more value-adding tasks.
Our computer-vision based solution is centered around the strengths of both AI and human experts. In model training, an unsupervised algorithm processes images of manufactured parts, pre-labels them as compliant or as anomalies and clusters them into several groups. An expert evaluates these groups (one image per homogeneous group rather than each component individually) and creates a labeled sample on which a supervised algorithm is trained. This results in a predictive model that is able to label any newly manufactured component as compliant or defective.
The resulting solution achieves several key benefits: Inspections are faster (<0.4s), more accurate (<500 PPM), and require fewer workers dedicated to manual control, decreasing total operational cost of one production line by 15%.