Automatic inspection of parts at factories using cameras and software is nothing new. But for a system to learn to spot a crack, a scratch or a bad weld seam, it usually needs to be shown thousands of labeled photographs of exactly such defects. Assembling such a dataset takes months, and for a small or medium-sized manufacturer the cost can prove prohibitive.
A group of researchers from the US and Europe, including mathematician Ying-Ju Chen of the University of Dayton (USA), tested a different approach. Their work is published in the Journal of Quality Technology. The scientists found that a model created by OpenAI is capable of finding defects on industrial parts if it is shown a relatively small number of examples of a specific flaw. The secret is that the model had already been trained on hundreds of millions of ordinary images, so it does not have to learn to recognize a defect from scratch.
"Automated visual inspection is not new, but many existing methods require large volumes of labeled data that are difficult and expensive to collect," Chen explained. "We wanted to find out how well a pretrained AI model can identify defects from a small number of examples and whether such an approach is applicable to different manufacturing inspection tasks."
The researchers tested the method on several inspection tasks. The results showed that a pretrained model can be adapted using only a moderate number of images. The paper also offers practical recommendations for applying and evaluating the method, including its strengths and limitations.
It is important to understand that this concerns laboratory tests, not a finished product that will go on sale tomorrow. The authors do not claim that the method fully replaces traditional preparation or guarantees error-free operation at any manufacturing facility. However, in their view, the approach could lower implementation costs and make automated inspection more accessible to manufacturers, including in the aerospace industry.