Optical systems for automated quality control have long been standard in numerous industries. These include simple color sensors and code readers as well as complex 3D multi-camera measuring systems. Conventional image analysis methods are used in many applications. However, when using AI algorithms, optical quality inspection systems in industrial assembly and production facilities require a large number of images for training. However, testable components or products are generally not available before the start of series production - and therefore no images are available. Effective training can therefore only begin with the start of production. Depending on the task and complexity, the final 'go' for the inspection process is sometimes delayed. If the number of units to be tested is small, the effort and costs per test part are comparatively high. This was also the case at the BMW plant in Dingolfing until July 2023: a test cell could only be put into operation once sufficient reference images of real vehicles were available - sometimes a Herculean task that took a considerable amount of time, depending on the scope of testing and features. This is why BMW, together with the image processing experts from VisionTools, pushed for a solution that would provide reliable image data even before the start of series production.

The test cell

Every minute, models of the new 4-series convertible, among others, enter the test cell via an assembly line. With the help of two cameras, the station analyzes fully automatically whether three rubber plugs are present or correctly inserted in the area of the two rear lights. The camera system immediately registers errors and irregularities of any kind. An employee receives a repair order via the display.

From CAD model to virtual training image in just a few steps: Synthetically generated camera images from CAD data for AI training for testing from component 1 in the real image processing process.

© Vision Tools

Only at first glance does the inspection cell implemented by VisionTools for checking plugs on the rear light well look like a normal station. But behind this is a fundamentally new approach that is tantamount to a paradigm shift in testing and inspection processes: based on the CAD data provided, the image processing solution generates any number of virtual, realistic mapping variants of the test field at the touch of a button before the start of series production. All characteristics that occur in the real test run and can influence the image result are included in the simulation. For example, position and location tolerances of the workpiece and the relevant image processing components such as the camera and lighting are taken into account. Surface properties such as material, color and reflectance can be randomized in the simulation to a much greater extent than would be the case in reality.
would be the case in reality. Properties such as optical distortion and the depth of field of the camera and lens are also taken into account. As the dimensions of the test station are generally known, cameras and lighting can also be positioned in the virtual space to optimize the results.