Inspection Technology Incorporates AI to Detect Defects in Real Time

Researchers at the Fraunhofer Institute for Material and Beam Technology IWS (Fraunhofer IWS) have developed a solution that uses AI and optical measurement technology to detect, classify, and visualize defects in real time, and report them to the plant carrying out the production. This process can capture 3D information of surfaces quickly and in high resolution. The measurement data is used to generate supplementary information in-line for ongoing production processes.

According to Christopher Taudt, group manager for surface metrology at the Fraunhofer Application Center for Optical Metrology and Surface Technologies AZOM, which is part of Fraunhofer IWS, the system, which is called SURFinpro, both detects defects and classifies them at the same time. Customers, Taudt said, receive information about the type of defect detected, as well as information about many other parameters — such as the defect’s density, geometric dimensions, and frequency.

Using artificial intelligence and optical measurement technology, SURFinpro detects, classifies, and visualizes defects in real time as the process is ongoing. Courtesy of Fraunhofer IWS.

The measurement system has been operating successfully in industry for over a year, analyzing a roll-to-roll process with substrate widths of 70 cm. To leverage further potential for optimization, Taudt and the SURFinpro team are now training the system within ongoing production using a defect catalog. As defects are reported, they are fed into a neural network, thereby refining the detection accuracy. The researchers use the measurement information to check if new defects occur or existing defects are modified, which requires a dynamic response of the system.

In addition to working to develop better neural networks that require less data, according to Taudt, the team is also developing new training strategies within ongoing operation. The researchers are working to adapt the technology to new fields of application, such as continuous fiber-composite manufacturing processes.

“Here, our partners are not only interested in avoiding near-surface defects, they also want the technology to be able to identify and assess components in multiple dimensions,” Taudt said. Another target group that the team envisages for the algorithms and defect-classification system is the semiconductor industry, for example in the production of flexible semiconductor materials.

Currently the Fraunhofer AZOM solution uses a maximum of four cameras. In a next step, the team plans to add additional camera systems. This would be beneficial regardless of the process being assessed, from fiber-composite processes involving very large components to traditional roll-to-roll processes, such as those used in the photovoltaic industry, for example.

Another key aspect for the scientists is the system speed. For fiber-reinforced plastics and textile processing, very short cycle times need to be particularly fast.

According to the researchers, one of the key characteristics of their system is its modularity. Due to a sophisticated modularization approach using efficient components, SURFinpro provides a wide variety of potential deployments and is easy to adapt. Many of the technologies used in the current system were developed as standalone components, though in a way that ensures they can also be implemented in various other contexts, Taudt said.