The solution is based on deep learning, a special form of machine learning that can operate with very unstructured and high-dimensional amounts of data such as with images. The team spent months training the artificial neural network with several million test images. The biggest challenges were on the one hand, the creation of a sufficiently large database, and on the other hand, the so-called labeling of the images. The team marked cracks in the sample images with pixel precision – the highest degree of accuracy was required. The effort was worth it because the neural network now learns independently from the examples and detects cracks even in new, previously unknown images. The database consists of several terabytes of test images from seven presses at Audi’s Ingolstadt plant and from several Volkswagen plants.
“Artificial intelligence and machine learning are key technologies for the future at Audi. With their help, we will continue to sustainably drive the digital transformation of the company,” emphasized Frank Loydl, Chief Information Officer (CIO) at AUDI AG. “In this cross-divisional project, we are jointly developing a production-ready solution that Audi will use exclusively in the company and which is unique in the market.” The software was mainly developed in-house, from the idea to the finished prototype. Since mid-2016, the innovation department of Audi IT has been working hand in hand with the Production Technology division of the Equipment and Metal Forming Technology Competence Center.
In the future, quality inspection using ML will replace the current optical crack detection with smart cameras. This involves a great deal of manual effort. Whether doors, engine hoods or fenders – the camera currently has to be reconfigured for every new component produced in the press shop. In addition, false detections regularly occur, since the simple algorithms of the image-processing program are highly dependent on ambient factors such as lighting conditions and surface properties.
In the future, it will be possible to apply the ML approach also for other visual quality inspections. If a sufficiently large number of labelled datasets are available, the system can also support paint shops or assembly shops, for example.