Johnson Electric is the world leader in electric micromotors for automotive and consumer applications. Johnson Electric produces millions of motors each day on high-speed automated production lines. Quality requirements are extremely strict and 100% of the product must be thoroughly tested prior to shipment. This testing can take up several production cycles, thus becoming a very costly element in production. In addition, sometimes new, previously unknown, error modes appear in production, which are hard to detect with classical rule-based algorithms. Machine learning can help speed up quality control, lower scrap rates, and avoid costly line stoppages. The main barrier to introducing machine learning in a cost-effective and saleable manner is that it is usually very hard to get a good reference set of labeled ground truth examples. Especially samples of defects are hard to find because of low intrinsic scrap rates and previously unknown defects. In this talk we explore the use of unsupervised learning for a key quality control application to address these issues and create a scalable machine learning solution for end of line quality control.