Virtual commissioning of AI-based camera systems

Optical systems for automatic quality check are standard in many industries. These systems range from simple colour sensors and code readers to complex 3D multi-camera measuring systems. Mostly, conventional evaluation algorithms are used for image analysis. In recent years, however, artificially intelligent methods, especially deep neural networks (CNNs - Convolutional Neural Network), have also become established.

However, these algorithms require a sufficiently high number of images, both positive and negative examples, for training. Otherwise, there is a risk that AI-based inspection systems will not deliver reliable or even incorrect inspection results. Since usually only a few sample parts or prototypes are available at the start of series production, the image data necessary for training the AI is also missing until then.

The solution to this problem was to develop a tool for generating virtual image data. As part of the "AI Innovation Competition Baden-Württemberg", the "SyDaVIS-AI" project of the consortium consisting of the Karlsruhe University of Applied Sciences - Technology and Economics, VisionTools Bildanalyse Systeme GmbH (Waghäusel) and Lensation GmbH (Karlsruhe) was awarded research funding by the state of Baden-Württemberg in 2020.

In the virtual environment, the machine components, the components to be inspected as well as cameras and lighting are combined to form a digital twin. By simulating the camera and lighting and their specific parameters, virtual camera images are generated with the help of synthetic image generation. Not only virtual positive examples, but also false or defective images can be simulated, all without a real workpiece. Recently, these image data have reached such a level of quality that they can now also be used to train AI algorithms.

Thanks to the use of synthetic data in training, the systems are now able to perform a reliable quality check from the first workpiece produced. For many application areas, valid inspection results are now generated as soon as series production or the manufacture of very small batches begins.

Optical inspection systems can thus be used effectively right from the start, without delays or erroneous results. Virtual commissioning with subsequent simulation and generation of image data is an efficient solution to accelerate the training process of AI algorithms and improve the quality of optical inspection systems.

These new innovative technologies enable a fast integration of the inspection systems and contribute significantly to optimising the production processes and ensuring the quality of the manufactured products.