Machine Learning – Not Without My Hardware

Fraunhofer IWU

Optical Inspection System with Hardware from mk

Intelligent software independently learns to detect, distinguish between and rate optical defects. The system described inspects small lacquered parts in a cycle. It also automates the quality assurance process in the production of small parts. The fact that the software can even see the faults is thanks to the use of the right hardware. The plant was designed by the Fraunhofer Institute for Machine Tools and Forming Technology (IWU) in Chemnitz, Germany. The frame for the cameras, light and computer as well as the conveyor technology were supplied by mk.

Large quantities in the highest possible quality – individual parts, for the automotive industry especially, have to meet high production requirements. An every increasing level of automation combined with intelligent machines offers opportunities to do justice to the task.

Intelligent machines. It sounds so simple. The term is often and widely used. But what does “intelligent machines” mean? A machine can only ever know what the programmer has taught it – so surely it can’t be that intelligent? Not true. Machines can learn independently, detect patterns and analyse a wide variety of data.

The Fraunhofer Institute for Machine Tools and Forming Technology (IWU) in Chemnitz, Germany has now developed a system that can detect defects on black-lacquered small parts for the automotive industry. Fully automatically. The software does not just detect established defect patterns. No, the software independently learns to detect, distinguish between and rate optical defects.

Intelligent Software

“We used our development environment Xeidana® to develop this system,” says Alexander Pierer, research associate in the automation and monitoring department. Alexander Pierer is responsible for research and development in the area of automation and monitoring of production technologies. He took on the leadership role in this project and helped to develop the software.

“One example of what the software that we implemented can do is combine the data from different or redundant sensor systems.” This sensor fusion gave them the opportunity to increase their range of identifiable defects and the rating reliability associated with that.

With the aid of structure-detection techniques, the software can find complex interrelationships in data pools and identify patterns. “And that’s how the machine learns,” explains Alexander Pierer. The parts can also then be classified as “okay” or “not okay” based on “soft criteria” that are similar to human perception.

High-Performance Meets High-Tech

The 19 cameras, which each take 35 images per second in full HD resolution, check up to 250 small parts for visually identifiable faults during the cycle. The workpiece carrier containing the small parts to be inspected is conveyed through the system below the cameras at 18 metres per minute. The cameras take 360 images in three to four seconds, which are then evaluated within a cycle time of 30 seconds. “To process the data, we use “massive parallel data processing”, involving 14 computer cores and the graphics processor. That allows us to equip the system with twice the number of cameras or more sensor principles without issue, which means we will be able to check other characteristics such as the colour tone or coating thickness in the future. We have incorporated sufficient reserves of mechanical installation space and software interfaces into the system concept,” explains Alexander Pierer. The cameras use deflectometry to identify unevenness at µm level. Scratches and inclusions are detected to a tenth of a millimetre.

Powerful LED lamps illuminate the products from three sides. To prevent motion blurs and the effects of extraneous light on the images, the products must be illuminated for a very brief time and with a high intensity: a flash with a range of 10 µs helps the cameras to take super-precise images.

Interplay Between Software and Hardware

In such projects, software and hardware must interplay seamlessly. “We still needed the conveyor technology and all the casings, camera holders, computer housing and a workstation for the system,” says Alexander Pierer. Pierer had positive experiences in earlier projects working with the profile and conveyor technology specialist Maschinenbau Kitz GmbH from Troisdorf, Germany. “We had been very happy with all our collaboration up to that point. And mk were once again the best equipped to meet our requirements in this project.”

This project primarily involved extremely flexible setting options for the cameras. To ensure the parts could be thoroughly inspected, the cameras had to be precise to within 1/10 mm and adjustable on three axes. The mk designers and IWU staff worked together to find an efficient and flexible solution in this regard that could accurately meet the requirements.

Alexander Pierer, Fraunhofer IWU, Chemnitz
We had been very happy with all our collaboration up to that point. And mk were once again the best equipped to meet our requirements in this project.

A Concept Brought to Life

Gabriel Jaramillo, the responsible designer at mk, managed to supply the right ideas: “We received a concept drawing from the customer showing the angles in which the cameras are set and how the whole solution might look. And we then built our design around that.” Dirk Hoffmann, designer at the Fraunhofer IWU, created this concept model: “Our collaboration with the mk designers worked extremely well,” says Dirk Hoffmann. “They expertly brought our preliminary design to life.”

Special requirements were also not an issue. For instance, the materials used had to be free from substances that interfere with the lacquering process and the design also had to be antistatic. Housings for the computer, brackets for the monitor and keyboard and a ZRF-P 2010 timing belt conveyor that served as a key component of the system, conveying the workpiece carrier through the entire plant, were all included in the scope of the project. The motor is connected directly to the drive shaft, reducing the amount of space and maintenance required. The drive position can be set anywhere along the entire length of the conveyor, which makes the conveyor extremely easy to integrate. A rotary encoder on the conveyor motor tells the software the exact position of the workpiece carrier on the conveyor to a hundredth of a millimetre.

The whole camera area had to be provided with a non-transparent casing. “We did not need a proper darkroom because the powerful short-term exposure can compensate for the effect of extraneous light to a great extent,” explains Dirk Hoffmann. “However, we did need the casing to be non-transparent, primarily to provide a glare shield for operators and protect the optics.” Furthermore, the plant had to be easy and quick to open for maintenance purposes.

Modular System

The mk plant could do all that. It did so thanks to the extensive modular mk system of profile technology and conveyor technology, which lets you implement even whole system solutions with ease.

“The job didn’t pose any major challenges for us,” says Gabriel Jaramillo. He was able to use or adapt much of mk’s standard repertoire. In turn, the use of standard components also has a favourable effect on the price. “The price-performance ratio impressed us,” says Alexander Pierer. He had finally found what he needed. “The system is running smoothly at our site, to our complete satisfaction.” The final integration of the overall system into the end customer’s facilities is still pending. The system will be integrated between the production/lacquering step and the step for separating defective parts. Thanks to the benefits of the mk profile system, no great difficulties in integrating the hardware are expected.

In the Future

For the end customer, the system means a seamless 100% control system for all its parts. Controls were previously performed by staff. The automated inspection should make quality assurance even more precise and quick. Staff capacities are freed up for other tasks that cannot be performed by machines, because despite all this machine intelligence, some human skills are too difficult or even impossible for a machine to learn.

Fraunhofer IWU itself is best able to develop software and task-specific hardware components. However, conveyor technology and mechanical plant assembly are not part of IWU’s core business. Alexander Pierer is sure there will be future collaboration with mk: “We will be happy to contact mk again for future projects. Creating the right surroundings for our software is just as important as the software itself.”

About Fraunhofer IWU

As a leading institute for resource-efficient production within the Fraunhofer-Gesellschaft, the Fraunhofer IWU ist a scientific research and development partner for the future industries of automobile and mechanical engineering. Since 25 years, the main focus of their work has been on application-oriented research and development in the field of production technology for the automotive and mechanical engineering sectors. Fraunhofer IWU not only develops intelligent production systems for the manufacturing of car body and powertrain components, but also optimizes their related forming and cutting manufacturing processes.

Fraunhofer Institute for Machine Tools and Forming Technology IWU
Reichenhainer Straße 88
09126 Chemnitz

Germany

 

Phone +49 371 5397 0

www.iwu.fraunhofer.de