To the human eye, the metal part appears perfect. However, thanks to the three-dimensional camera images, the automated inspection system detects tiny deviations that exceed the permitted limits. This makes it clear that the component does not meet the quality requirements and must be rejected. Such deviations in the micrometer range can now be detected very accurately thanks to increasingly powerful image processing technologies.

In the course of Industry 4.0, the demands on image processing methods are increasing even further: markets and market requirements are changing rapidly, customers are demanding increasingly individualized products with consistently high quality standards. This has consequences for production: product life cycles are shortening and batch sizes are becoming smaller. In addition, there is greater time pressure due to automated and extremely fast production processes, which make parallel quality controls necessary - with maximum precision.

Suppliers of image processing systems have adapted to the accelerated production processes in recent years. Image processing has become faster and more user-friendly and benefits from high-speed cameras with high resolution and powerful interfaces, for example. But that alone is no longer enough: to ensure smooth production processes in the age of Industry 4.0, machines need to be easily and flexibly integrated into production lines. In addition, new forms of collaboration are needed that combine the precision and endurance of the machine with the knowledge and problem-solving skills of humans.

Machines can learn too

Users of image processing systems face two major challenges, particularly with regard to the required flexibility: On the one hand, a large number of training runs are often necessary before the system can distinguish defective from non-defective test pieces. On the other hand, adapting existing systems to new tasks often involves a great deal of effort.

Completely new standards are set by systems that are able to communicate with their environment and learn from it or directly from humans - keyword 'learning image processing'. This refers to algorithms and programs for image analysis and pattern recognition that can be adapted to new tasks without time-consuming training or can even learn them themselves.

One example of such a learning image processing solution is the 'Apas inspector' from Bosch. This automatic inspection system includes a flexible platform with a planar table, a high-resolution high-speed camera and a touchpad. The inspection system inspects small to medium-sized batches with matt or glossy surfaces. The touchpad enables intuitive operation, while highly developed 3D imaging processes deliver reliable, precise results even under harsh production conditions. Thanks to learning image processing, continuous adjustment of the inspection parameters is possible.

Simple re-training

What does this look like in practice? Using training images, the test specimens can be classified into 'good parts' and 'bad parts', for example. To do this, an employee uses the Apas inspector to take pictures of several test specimens and mark the 'good' and 'bad' areas. The system analyzes and saves the information for future inspection runs. The inspection system can be retrained at any time with additional images. This allows the user to continuously improve the system's detection performance or adapt it to changing conditions. The number of classes and features according to which the learning image processing system distinguishes the test specimens can be freely selected for each inspection task and can also be changed at a later date.

The images captured with the 'Apas inspector' can be classified into 'good parts' and 'bad parts'.

© Robert Bosch

This is what happened at a Bosch plant when testing the functional surface of a sealing ring for a hydraulic actuating unit of an automatic transmission: if the sealing rings show the smallest grooves or burrs after the last machining, this can result in a functional failure. As a result, the automatic transmission does not shift, but goes into emergency mode - the smallest causes can have correspondingly large consequences. A 100% inspection with the Apas inspector provides a remedy here: the sealing rings rotate under a line scan camera. Using photometric stereo, a 3D representation of the circumferential surface is generated, which highlights the smallest surface deviations - potential leaks. It is important to differentiate between regular machining marks and defects.

This is where adaptive image processing comes into play: the production experts can mark the unproblematic, regular surface marks as 'good' and the burrs and grooves as 'bad' in the captured images. With the help of this intuitive learning process, 90% of all defects could be reliably identified with just a few training images. After further retraining during commissioning on site, the Apas inspector was finally put into three-shift operation. An optical 100% inspection was installed, with only around four weeks between inspection of the sample parts and acceptance in production.

Man as teacher

Employees can continuously improve the system's detection performance - by becoming teachers. Based on their experience, an employee knows which scratches detected are relevant to quality and which are not critical. They use the touchpad to mark 'good' - i.e. qualitatively flawless - and 'bad' areas - those with unwanted scratches, holes or flaws.

Smooth production processes in Industry 4.0 require machines that can be easily and flexibly integrated into production lines.

© Robert Bosch

This teaches the machine to distinguish between 'good parts' and 'bad parts'. However, the employee can also let his 'pupil' go through supposed errors and thus teach him which parameters are actually relevant to quality for a particular task. As the system stores all the information for future inspection runs, its 'memory capacity' is constantly growing. For further inspection tasks - for example color inspections or surface measurements - the employee can freely select the number of features according to which the learning image processing distinguishes the test specimens and subsequently adapt them to changing conditions. In this way, the employee 'translates' part of his wealth of experience for the machine. He does not need extensive programming knowledge for this, as even an inexperienced employee can show the image processing system the relevant features and thus 'train' it to achieve higher performance and precision.

Author:
Wolfgang Pomrehn is Product Manager APAS Assistance Systems at Robert Bosch in Stuttgart.