Artificial Intelligence (AI) and Machine Vision can make a significant contribution to increasing product quality in the manufacturing environment. Kontron has not only put the benefits of AI into practice in customer projects, but is now also using it in its own visual quality inspection.
Quality control in discrete manufacturing is particularly complex when dealing with small batches and a large number of variants, as is the case at Kontron. Machine Vision and Artificial Intelligence are playing an increasingly important role here in order to improve product consistency and quality. AI-based object recognition systems can detect tiny imperfections and discrepancies that would be almost impossible to detect during a human inspection. And unlike the human eye, cameras also never get tired – which also helps to reduce error rates.
Automating highly complex quality inspections
Kontron`s customers order tailor-made systems with specific ports and connectors, often with optional features and individual mechanical configurations. To ensure that the systems have been configured and assembled correctly, each product must be compared to its specification – a time-consuming and therefore costly task that the AI solution was designed to address.
In collaboration with Intel®, a visual AI-based inspection solution was created based on Intel® Edge Insights for Industrial (EII), running on the OpenVINO™ development architecture. The solution benefited from the fact that the existing PCs with 11th generation Intel® Core™ processors could be used continually without any additional special hardware – such as discrete graphics processors or accelerators. A tight time frame was set for the implementation, in which the ongoing production process had not to be interrupted. Intel® EII and the OpenVINO™ toolkit provided the crucial functions for this.
Close to the edge for greater efficiency
To increase speed and cost efficiency, the complex computing processes take place directly at the edge, close to the data source. Video data acquisition, storage and analysis were therefore handled in a single software package. To train the model, six installed cameras captured the streaming image data of the inspection area from different angles. This provided a multi-dimensional view of each product.
The test with commercially available USB video cameras for the consumer market yielded a surprising result. The original plan was to use expensive, specialized high-end cameras and additional hardware for image recognition. However, the USB cameras delivered sharper images with better detail resolution – presumably due to the autofocus function – and were much easier to network: a real benefit to the project.
Augmentation helps with object recognition
Experience shows that the more labeled comparison images are available for an AI algorithm, the better the results. Therefore, one focus has been to generate as many additional images as possible to increase the accuracy of object recognition. For this purpose, techniques such as augmentation were used, e.g. rotating the images by a few degrees, using different brightness values, cropping, filling and mirroring techniques.
More than 12,000 different images and over 30 labeled objects were used to train the algorithm, which now compares the currently captured images with what it has learned in real time and evaluates the quality. In addition, learning models and methods offered by the Intel® OpenVINO™ Model Zoo, such as transfer learning, were experimented with. Good results were achieved with the easy adaptation of the algorithm for comparable components, such as different screws. In the future, the solution will also be flexibly adaptable to new products using new training data.
The new inspection solution has quickly delivered tangible benefits. The project has shown that with the help of AI technology and a relatively simple setup, it is possible to automate visual quality control in small batch production cost-effectively and quickly.
Explore the Kontron Group at: https://www.kontron.com/en