Completely under control

Why did the AI specialist choose the uEye XC camera from IDS?

"In addition to the interface to the self-developed cam controller and the compact size, the range of functions of the sensors and, of course, the price due to the number of cameras required, also played a role," says Eike Humpert from Vivaldi's Business Development department. "The decisive factor was that the autofocus camera can capture several surfaces at different distances from the camera." This simplifies the testing process enormously and ensures greater efficiency and flexibility. This is a crucial factor when, for example, different component variants need to be tested. With the macro attachment lens, the minimum object distance of the camera can also be shortened very easily.

The 13.10 megapixel CMOS sensor AR1335 from onsemi, with which the compact uEye+ XC is equipped, is also shown here. The rolling shutter color sensor (4200 x 3120 px, pixel size 1.1µm) delivers 20.0 frames per second and ensures detailed images. "The BSI pixel technology enables the sensor to deliver consistently precise and reliable results, even in changing light conditions," explains Jürgen Hejna, Product Owner 2D cameras at IDS. Its excellent low-light performance and minimal pixel noise are not the only reasons why it is particularly suitable for this application. "The 24x digital zoom, combined with auto white balance and precise color correction, contributes significantly to the success of this demanding test procedure," emphasizes Jürgen Hejna. With dimensions of only 32 x 61 x 19 millimeters, the camera is compact, sturdy despite its light weight and absolutely suitable for industrial use. The 2D camera expert mentions another plus point: "The components used here are characterized by long availability, one of the most important requirements for solutions for the automotive industry."

AI-supported analyses and synchronous camera control via edge server

The simultaneous image processing task is complex and the challenges are diverse. Aluminum parts often have very shiny surfaces that reflect light and therefore have a strong influence on images. These reflections make it difficult to reliably detect defects such as scratches, cold runs, chatter marks or surface irregularities, as they are often obscured or distorted by light reflections. In addition, it is often the case with aluminum parts that untreated surfaces look different every time. Areas where no processing has taken place show different structures and color nuances, but these should not be confused with real defects. AI-based image processing must therefore be able to distinguish precisely between the normal material structure and actual raw areas in surfaces to be processed that are considered defects.

Image processing is also made more difficult by the ambient conditions in production halls. Factors such as air quality, particularly due to a mixture of cooling lubricants, can affect the clarity and sharpness of the images. In addition, there are temperature fluctuations and vibrations caused both by the mechanics in the box itself and by other machines in the vicinity. These vibrations can lead to blurred images or positional deviations, which makes fault detection even more difficult.

Precise coordination of the image processing systems was necessary to meet these challenges. Lighting, camera positioning and image evaluation algorithms have been adapted so that reflections are minimized and material differences are interpreted correctly. At the same time, the system is robust enough to compensate for interference from environmental factors and deliver consistently reliable results. Defective components must be able to be replaced quickly and easily by the customer's own maintenance department.

"Image processing is carried out via an optimized infrastructure. Up to eight cameras can be connected per USB hub, which enables simultaneous recording of several inspection areas. A maximum of two of these USB hubs are connected to a self-developed 'cam controller', which bundles and processes the image signals. Several cam controllers are networked with the edge server. This server not only calculates the AI-supported analyses, but also communicates directly with the system controller, which coordinates the triggering of the cameras and ensures synchronous recording."

This has two advantages: Firstly, the final inspector is introduced to the digital final inspection step by step and becomes accustomed to working hand in hand with Vivaldi. Secondly, the labeled product images with error marking and classification are used to train the AI. In addition, existing product images are used for the training, e.g. recognized defective parts from the past, which are documented as part of the complaints process.

"By combining manual and automated surface inspection, we can achieve an error rate of less than one percent, even with large quantities," emphasizes Eike Humpert. At the same time, this approach allows the AI to be continuously trained. The user has constant control over the extent to which the AI checks and controls.

Another outstanding feature is the automatic image storage for each individual serial number, which ensures complete traceability of the process and quality data. This means that test results can be traced and documented at any time - a decisive factor for quality assurance in one of the most demanding industries in the world. "Humans must continue to have the last word," says Uwe Siegwart. "Everything has to meet the strict QA guidelines of the OEMs. The complete value-added life cycle of each component must be documented in an audit-proof manner. That means which part was on which system when and was tested with which test equipment and which results." The Vivaldi system guarantees all of this:

"Our technology is flexible and the platform is scalable. It is optimally adapted to the latest technology and can be seamlessly integrated into a wide range of manufacturing processes," explains Uwe Siegwart. "Thanks to the scalability of the solution, companies from a wide range of industries can adapt their quality control to individual requirements." Intelligent data processing of process and quality data enables sources of error to be identified at an early stage, leading to continuous process improvement in the core process of manufacturing companies.

"Digitized quality data acquisition in real time and linking with other production parameters not only increases efficiency, but also significantly reduces rejects and rework. Errors are minimized, costs are reduced and production processes are optimized." The system has already paid off in the Czech Republic: The number of complaints per year has been reduced by two thirds and the number of undetected surface defects has even been reduced to zero.