This is #newtk: Artificial neural networks for quieter steering gears
Engineering | Smart factory | trends of technology | All over the world, our colleagues at thyssenkrupp are working to safeguard the future of our company by making it more flexible, more efficient, and more high-performance. In our new #newtk strategy, we are putting performance center stage – and innovate our business with smart ideas. One of them: using artificial neural networks to determine the quality of ball screws much more precisely than before.
When talking about the acoustic experience of driving, the first thing most people think of is engine noise. But there is another element that has a major impact on the driving experience – the steering gear, which has a significant influence on background noise in the vehicle. One reason for this is that the steering gear transmits the movements of the wheels into the interior of the vehicle through vibrations. But noises that are deemed unpleasant are supposed to be as quiet as possible, whereas warning signals have to be heard by the driver.
So in the production process, completed steering gears are subjected to extensive final inspections, during which – among other things – their acoustic properties are measured precisely. However, these gears are made up of dozens of moving and rigid parts, many of which impact the overall quality. One that is especially important here is the ball screw. It transmits the steering movement to the axle like a worm gear and makes a significant contribution to the vibro-acoustic properties of the whole system.
Steering gear acoustics: No crystal ball for road behaviour
The properties of the ball screw are therefore examined first during the production process. It is rotated on a test stand and the resulting vibration frequencies measured. If their amplitude within specific areas of the observed spectrum exceeds the limits set by acoustics experts, the ball screw is classed as a reject. If, on the other hand, it passes the test, it can be fitted in the steering gear – which should then, ideally, also get through the acoustic test without a hitch.
But in practice this is not so simple, as a ball screw that tests “OK” can lead to a steering gear that fails the test. There are several reasons for this. First, it is hard to establish a correlation between the quality of the individual components and the steering gear as a whole, meaning that statistical methods have to be used for this. Moreover, the limits for the amplitude of the vibration frequencies are ascertained subjectively through test drives using pre-production components. And last, the final inspection is performed on the assumption that if a single acoustic limit is exceeded just once, this points to a substandard component. However, this can lead to “false rejects.”
Acoustic testing of steering gears: Lost in an ocean of data
In addition, there is another challenge, says Paul Alexandru Bucur, Project Manager Advanced Analytics at thyssenkrupp in Eschen, Liechtenstein: “Our acoustic checks produce acoustic recordings of around 6 to 12 seconds, from which our acoustic experts can draw conclusions about the quality of the components. In practice, however, such amounts of data from several million data points per recording are unmanageable for one person, and our acoustic experts just can’t listen to steering gears for eight hours a day.”
AI support: The artificial neural network listens
Together with its team of experts in Eschen, Bucur has therefore developed a method in which an artificial neural network based on special rules automatically recognizes common patterns in the vibro-acoustic behavior of ball screws and steering gears and thus achieves more reliable test results. However, the core idea of the project is that it is not the human experts who establish the required rules, but the neural network that automatically learns them.
“To do this, we trained a neural network at first,” explains Bucur. “This means that we ‘feed’ it with reference data sets from steering gears and ball screws for which the quality can vary, but is generally known. In the second step, the AI recognizes correlations within the data. This allows conclusions to be drawn as to whether the quality of the steering gear will be good or bad with a particular built-in ball screw.
Within this process, the neural network examines not just individual parts of the frequency spectrum but all of them. In this way, it is able to identify complex connections between the properties of the components. The insights gained as a result of this then help in testing the ball screw first, as usual, so as to draw conclusions about the quality of the subsequent steering gear.
Neural network leads to new insights into product quality
The new process has proven successful. The artificial neural network has been able to detect data that were not covered by the previous quality measurement procedure. It is these very data that have also yielded information about the quality of the ball screw, and this has enabled the focus on the test stand to be adjusted accordingly. The results of the learning process in the network were not abstract patterns but can be traced back to physical effects in the production process.
The new process is also very important to thyssenkrupp in business terms, for the company produces large numbers of ball screws every year at its Schönebeck site. From there, they are distributed to other sites in Europe and all over the world, where they are fitted in steering gears. lt takes two to three days to transport them within Europe and six weeks to deliver them to Central America and Asia. This is why it is especially important that these components should not cause any problems when they are finally installed in the steering gears.
Introduction marks continuous change process
Bucur, a trained mathematician and former software developer, regards the introduction of the neural network as the starting point for a long-term process involving numerous organizational changes. “This is not just about technical aspects. In the past, we were able to see immediately, for example, if certain frequency ranges caused problems, based on the clearly defined rules set up by our colleagues. With the neural network, however, there is no such form of comprehensible rules, because the patterns learned independently by the network are very difficult for humans to interpret.”
Since it would be very dangerous to trust the neural network blindly, the team has successfully visualized the complex patterns. “This enables us and our acoustic experts to check whether the content learned from the neural network is physically relevant,” said Bucur. In addition, the expert points out that what has been learned must be constantly checked. Not only do the samples themselves and the products change over time – the internal production processes and the quality of the materials used are not identical throughout.
Test pilot makes steering gear more reliable and less expensive
Nevertheless, the added value for acoustic testing more than outweighs the organizational challenges. The artificial neural network is still in the testing phase, but positive effects are already noticeable, says Bucur: “The decisive advantage of the neural network is that we can now detect at an early stage – for example during production – whether individual components such as ball screws do not meet our acoustic quality requirements. And not only after the entire steering gear has been assembled. This saves time, material – and ultimately costs.”
And so, the artificial neural network contributes to greater reliability in the assessment of ball screws and, consequently, of the entire steering gear system. This has resulted in a significant cost saving. So it is no surprise that other areas are also running projects using artificial neural networks, for their greater predictive power means artificial neural networks can also be used, for example, to shorten the testing period for final inspections in industrial production.