Machines Learn How to Learn from Steinbeis Experts

Artificial intelligence in mechanical engineering: machines become capable of learning

Controlling modern machines and robots has become so complex that classic control technology and other programming methods are reaching the boundaries of technical feasibility. Artificial Intelligence and Data Security (German: KIDS), a Ravensburg-based Steinbeis Transfer Center, is entering new territory in this field: For a number of years, Professor Dr. Wolfgang Ertel and his team have been looking at machine learning related to control, diagnostics and optimization issues in mechanical engineering. As can be seen from two projects conducted on behalf of Festo in Esslingen, their work has evidently been highly successful.

The LearningGripper is a grasping tool with four fingers designed as an abstract representation of the human hand. The four fingers on the gripper are moved pneumatically by twelve low-pressure bellow actuators. What is special about the bionic gripper is that it can learn. This is because it uses learning algorithms instead of highly complex programs.

Thanks to machine learning (a field of artificial intelligence), the gripper is able to teach itself complex tasks, such as grabbing and positioning a ball. For example it has to rotate a ball such that at the end a given side of the ball points upwards. The gripper uses reinforcement learning to learn this behaviour. While Festo was developing the gripper hardware, the KIDS Steinbeis Transfer Center invested exactly one year in the learning algorithms, which were successfully implemented and then presented at the Hanover Trade Show.

Based on these principles, self-learning (adaptive) systems like the LearningGripper could be integrated into production lines in the future and then optimize their behavior by themselves. The crucial advantage machine learning has over classic process controls is that the learning algorithm does not need a mathematical model for the hardware – in this example: the gripper. This is similar to how we learn complex sequences of movements as human beings, without having to study mathematics or mechanical engineering. Just like our brains, the learning algorithm generates its own model of the task it needs to learn.

Now that the gripper project has demonstrated the successful application of machine learning with a prototype, a second project was started to solve a more specific problem for Festo involving pneumatic machinery. The project aim was to continually gauge the flow rate of compressed air on an automated pneumatic device in order to detect changes in power consumption, irregularities and faults. This can be as a result of things like compressed air leaks or other defects in the pneumatic system.

The device to be monitored was treated as a black box, since the diagnostic module that was to be developed also had to be usable on any other kind of pneumatic device. It had to be possible to connect the diagnostic module to devices without major effort or expense to allow operators to carry out monitoring or function diagnostics. Furthermore, the module would have to “get to know” equipment in a learning phase of a few hours in order to be able to classify deviations from normal operating conditions as errors. To do this, machine learning algorithms were used that can detect deviations from the typical shape of the flow rate curve.

A particular challenge with the project was a condition laid down by Festo: during the learning phase, the device had to operate only under normal conditions. As a result, the learning algorithm had no access to faulty operating conditions of the device during the learning phase but still had to be able to distinguish with certainty between normal opera- tions and faulty conditions afterwards. To solve this problem, the experts used a modified version of a method called “one-class nearest neighbor.” The method was introduced to a prototype and tested very successfully with flow rate data on several machines – i.e. with an extremely low error rate.

It is often not possible to fit machinery with a sufficient number of sensors due to cost or technical reasons. This can result in major problems with model-based diagnostic methods. The new diagnostic module only requires one flow rate sensor, and, thanks to the machine learning process, it can extract relevant information from the flow rate chart.

The project is also interesting given developments regarding Industry 4.0, on the cusp of an era of autonomous machines. With modern mechanical engineering it is now possible to design extremely complex machines, and powerful sensors, drives and control units are entering the field from mechatronics and electrical engineering. Simultaneously, over the past 20 years, developments in artificial intelligence have resulted in learning algorithms that can be put to highly practical use, and, just like the task that was solved in this case, many areas are simply waiting for this to happen. With this project, two matching partners worked together, culminating in a pleasing result. But it was soon clear from the many meetings and discussions that the approaches adopted in mechanical engineering and artificial intelligence are a long way apart. The project was only possible in this constellation because of the strong interest from Festo in new methods from different specialist areas, and because the scientists at the KIDS Transfer Center were motivated by their enthusiasm for the practical challenges encountered every day in automated manufacturing.

The solution that was developed as part of this project is universally applicable and the sensors were extremely basic, meaning it could be applied to many other diagnostic tasks of a technical or non-technical nature. For example, this potential innovation could be applied to the automatic self-diagnosis of household instruments, electric motors and gasoline engines. Also, it could be used in different ways for medical diagnostic purposes or for home security monitoring.


Prof. Dr. Wolfgang Ertel is director of the Steinbeis Transfer Center for Artificial Intelligence and Data Security (KIDS), which is based at Ravensburg-Weingarten University of Applied Sciences. The center offers its clients R&D services as well as consulting services relating to the field of machine learning used in diagnostics, prognosis, classification and data security.

Professor Dr. Wolfgang Ertel
Steinbeis Transfer Center Artificial Intelligence and Data Security (Ravensburg)

Share this page