Improving the accuracy of mechatronic systems by learning from data using artificial intelligence
Johan Kon defended his PhD thesis at the Department of Mechanical Engineering on June 18th.
Machines that operate with extreme precision are used extensively around the world, from the machines that produce microchips for electronics to medical imaging devices that make minimally invasive surgery possible. These machines operate with an accuracy ranging from the width of a hair to one-thousandth of that. As these kinds of machines become increasingly complex and face stricter demands on speed and accuracy, traditional algorithms for controlling their motion are no longer sufficient. Within his PhD research, Johan Kon focused on improving the precision of these machines by learning their behavior from data using artificial intelligence.
To be specific, Johan Kon used neural networks for his research. These are algorithms inspired by how human brains process information. Unlike traditional algorithms, which follow explicitly programmed rules, they learn patterns from data. For example, neural-network-based language models such as ChatGPT can learn to generate text that looks as if it was written by a human by analyzing large amounts of written language. Similarly, neural networks have the potential to learn how machines behave by identifying patterns in motor and sensor data.
From the digital to the physical world
Despite the success of neural networks in the digital world in fields such as language processing and image generation, their potential remains underutilized in the physical world because of concerns about safety and reliability. When a chatbot makes a small error, the worst outcome might be a badly rhyming poem or nonexistent restaurant suggestion. In contrast, when neural networks interact with machines in the physical world, even small errors can lead to hardware damage or safety risks. That鈥檚 why it鈥檚 essential to incorporate system knowledge and theoretical guarantees when using neural networks to learn patterns from data.
Johan Kon developed a method that allows neural networks to augment existing physics-based models rather than replace them. This ensures a baseline level of performance and enables monitoring of the neural network鈥檚 contribution, or even disabling it when necessary. Additionally, he has designed a technique that constrains the neural network over time, so the system continues to operate safely, irrespective of what the network has learned from data.
Real-world applications
The developed methods form a framework for the safe and reliable use of neural networks in high-precision applications. In this research, the methods were applied to improve the precision of a medical X-ray system by learning the effects of cable and friction forces from data. This is too time-consuming to model using traditional methods. As a result, the system achieved higher accuracy compared to traditional methods. In another example, the methods were used to enhance the performance of a stage system, a critical component in semiconductor manufacturing. By learning how the stage deforms over time, neural networks can improve precision, while safe operation is guaranteed through the developed methods. These results demonstrate how neural networks can be safely and effectively used for high-tech applications, opening the door to faster and more accurate systems in the future.
Title of PhD thesis: Supervisors: Prof. Tom Oomen, Prof. Marcel Heertjes and Prof. Roland T贸th.