This project extends existing optical quality control methods by machine learning techniques to reduce the pseudo error rate in automatic optical inspection (AOI). As a result, fewer manual follow-up checks are required, and workers are relieved of monotonous tasks.
Many manufacturers use AOI as a means of quality control in SMD production (Surface Mounted Devices). These systems use conventional image processing algorithms, which makes them high-maintenance and much less dependable where product variants are concerned. The automated inspection process therefore generates many pseudo errors, with the consequence that all hits need to be checked by hand.
In the course of the project, a machine learning technique was developed that can distinguish between pseudo errors and real errors in an AOI. The technique is highly robust in the face of varying components. This greatly reduces the need for manual rechecks and eliminates these monotonous tasks. To better understand how the new technology works, the algorithm was implemented in a demonstrator that independently takes photos of defective components and thus imitates the AOI process (see video). The software is easily updated by simply adding new components to the training data set and carrying out a short re-training procedure.
Sponsored by the State of Baden-Württemberg, Ministry of Economics, Labor and Tourism