Skip to content Skip to sidebar Skip to footer

The artificial eye for plastics

Source: Der Standard as of 26-02-2020

Automated quality control plays a major role in industrial production. Leoben researchers use artificial intelligence to be able to flexibly detect surface defects in workpieces.

Two important applications of artificial intelligence (AI) in industry are quality control and predictive maintenance. In one case, it is automatically determined whether a component actually meets the required specifications, the other is to react to irregularities before the plant is shut down according to the motto: the spare part is already just before the Swap breaks. A development of the polymer competence center PCCL in cooperation with the Montanuniversität Leoben shows how one can combine both principles in order to optimize production in a kind of “predictive quality control”. The quality of the manufactured workpieces is not only continuously checked. The smallest deviations that are registered are used to change the parameters of the machine and to improve the following workpieces. Dieter P. Gruber is working on his research group “Robot Vision and artificial intelligence at PCCL has been automating surface inspection of components for more than ten years. “It’s about developing an artificial eye that can detect visual disturbances in 3D components even if these surfaces have variable structures and patterns similar to a wood grain or a textile fabric,” the scientist says.

Scratches, streaks, holes Specifically, it concerns plastic cladding inside cars that have a decorative pattern. “Changes and distortions of the pattern and different reflection behaviour of the surfaces are quite permissible here, but not scratches, streaks, holes or so-called incident points, in which the stabilizing structures on the back of the components on the front,” Gruber explains. “The system must be flexible enough to allow for permissible changes and defects in the surface structure.” The test system developed in the research group usually uses multiple camera sensors and light sources. The resulting data material must also be able to characterize shiny and reflective surfaces well often a hurdle in automated quality control. In order to train a neural network to detect the defects, runs with series of test parts are made first. On the one hand, different classes of surface disturbances are presented to the system. On the other hand, you can work with good parts to let the AI “learn” the appearance of flawless specimens. After a general training that defines good and defective parts, the system is optimized with further strategies, such as specific post-training. As part of a production plant, the system is intended not only to perform optical faults, but also to shape changes such as the mentioned incident points. “The molded parts change imperceptibly when factors such as pressure or temperature are not optimal during production,” Gruber explains. “That’s why we measure up to the micrometer range how much the current quality of each component deviates from an optimum defined as a target. This measurement becomes a feedback for the production of the following parts.” Balancing deformation For example, there could be a condition that a surface must never have a deformation deeper than 20 microns, as it would become visible as a fault. If a series of workpieces has an increase in the value towards 20, the system can react automatically and, for example, increase a process temperature to counteract the trend and to achieve lower values again. Previous procedures, in which a whole batch of workpieces are first produced and tested in order to adjust a machine, are omitted. Committee is avoided. For Gruber, however, fault detection in surfaces is not the end of the flagpole. In his research group, he also uses AI algorithms to detect the feel of a plastic object more specifically, how a surface feels when touched. “A surface can be velvety, leathery or ‘frogy’. I have made it my mission to systematize this sense of touch,” says the scientist.
In experiments, the contact information of an artificial finger is measured on different surfaces with special sensors, the data is evaluated using the neural network and associated with parameters of production. In this way, just like the optics or robustness, the touch of a plastic workpiece could be better designed in the future.
Show CommentsClose Comments

Leave a comment

News ORS © 2020. All Rights Reserved.