For Efficient Production Transforming Quality Assurance with AI

Related Vendors

Fraunhofer Institut für Graphischen Datenverarbeitung

FAULHABER Drive Systems

Taiwan Machine Tool & Accessory Builders' Association (TMBA)

Wuxi InfiMotion Technology Co., Ltd.

Optimize production processes and increase efficiency—more and more companies are using artificial intelligence for this. But training is time-consuming. New methods are now paving the way for economical and accessible AI-based quality assurance.

Generative AI is already enriching the daily lives of many people, but industrial companies face a challenge in implementing this technology: for AI training, they need not only images of flawless products (short: OK data) but also hundreds of images of defective items (short: NOK data). What is generally advantageous becomes a hurdle in this case since production should ideally generate as few defective pieces as possible. Approaches like training solely with OK data and the synthetic generation of training data based on CAD data can solve the challenge.

Generate Artificial Training Images with CAD Data

At the beginning of a manufacturing process, there are no real photos yet—neither OK nor NOK data. The Fraunhofer Institute for Computer Graphics Research IGD is researching methods to generate images of three-dimensional models from CAD data using virtual cameras from various perspectives and orientations. They then virtually apply different materials and a variety of backgrounds to the component. "This allows for the generation of numerous images and the creation of training databases within a very short time, without ever having to add real photos," explains Holger Graf, head of the Virtual and Augmented Reality department at Fraunhofer IGD.

Gallery

The inspection system has never seen the real setup and product configuration before in operation, yet it can still recognize, classify the object, and estimate its position. Additionally, this approach shortens the reconfiguration time of the inspection system for any variant. Fraunhofer IGD developed the technology specifically for assembly or construction state control in automotive and commercial vehicle manufacturing, as well as in tool manufacturing. Another example shows the application of the solution in the production of airbag igniters. Here, automated optical quality control is particularly important: The end product is highly safety-relevant and cannot be fully tested—the airbags cannot be used again after deployment.

Optimal Classification Using only OK Data

In other applications, quality inspection cannot be based on CAD data. This is because they are either not available, or the product is to be assessed not in its original state but by its appearance after a stress test. As a result, companies need to train their AI systems with real data. To avoid relying on a multitude of NOK data, Ulrich Krispel and his team from Fraunhofer Austria developed a solution that learns solely from OK data. These are qualified as "in order" in terms of production, meaning they do not need reworking or sorting out. The method thus learns a variation of normality and eventually also recognizes deviations from it. This way, the AI can also find previously unseen errors. The known problem with training AI classically using NOK data is that it cannot react appropriately when confronted with an image that falls outside the known error classes. The reason is that the AI is only trained to recognize and classify known errors.

The Fraunhofer solution is based on transfer learning, which involves pre-trained neural networks developed and published for research. These networks have already learned which areas in the image to focus on for classification. The scientists adapt the model to the application using statistical methods. "Pre-trained neural networks make it possible to keep the training effort as low as possible," explains Krispel. The researchers also identify the appropriate model for each application, aiming to produce optimal results while also making quick decisions.

In addition to the shortest possible training time, a process-capable classification time is also crucial. After all, the inspection system should not disrupt production processes but rather enable continuous and uninterrupted monitoring of product quality in real-time. This way, errors can be detected and corrected early on.

A few NOK images are sufficient for evaluating the model. The AI marks deviations from normality with colors—blue for slight deviations, red for defective areas in the image.

Supporting Manual Quality Assurance

Whether it's training exclusively with OK data or synthetic training data, AI-based quality control contributes to the overarching goal of economical production. "Our experience shows that companies are curious and want to take advantage of artificial intelligence. We assist SMEs as well as corporations in finding a solution tailored to them, as every product has its own peculiarities and each production environment has different requirements," emphasizes Holger Graf. The Fraunhofer IGD also supports interested companies in terms of imaging systems and technical equipment.

With the two approaches presented, the effort for users can be significantly reduced without any loss in reliability in the decision between OK and NOK. By automating quality inspections with AI, companies ultimately reduce the need for manual inspection and minimize human errors, which in turn decreases scrap and rework costs. As a supplement, AI can thus also alleviate the shortage of skilled workers in quality assurance.

*Prof. Dr.-Ing. André Stork is the automotive industry leader at Fraunhofer IGD.

Machine Authentication

Machine Operator Authentication - Why knowing "who" matters