Revolutionizing dynamical processes modelling with AI
Gerben Beintema defended his PhD thesis at the Department of Electrical Engineering on March 21st.
![[Translate to English:]](https://assets.w3.tue.nl/w/fileadmin/_processed_/e/1/csm_Beintema_Gerben_EE_PO_VH_4007%20%281%29_4b5dad82e5.jpg)
In his PhD research Gerben Beintema has developed Artificial intelligence (AI) methods to better model highly complex dynamical processes. His AI-based modeling methods contribute to enhanced resource- and cost-effectiveness, and performance of, for example, water and energy distribution systems, chip manufacturing machines, health care, fluid systems, traffic, and many more.
The developed AI-based enables one to obtain highly accurate dynamical models of complex (nonlinear) processes directly from measurement data in an automated manner. This is enabled by using artificial neural networks to express the dynamical models and the use of thoughtful learning algorithms.
Quick dynamic learning
A key insight in Beintema’s work is that it is sufficient and often advantageous to consider only short subsections of the available measurement data. This can be thought of as giving the AI not the entire ‘book’ of data to learn the model, but only short paragraphs that contain the key information that enables quick dynamic learning. This works since the AI-based learning algorithm can get confused (unstable optimization) when considering the entire book at once whereas learning from paragraphs eases the learning algorithm.
However, since these subsections of measurements are relatively short, additional context is required for the rest of the measurements. We prove that the context can be effectively summarized with an encoder function that is co-estimated with the model. This enables AI to learn highly effectively only using limited computational resources.
Spatial awareness
These innovations apply to a large set of dynamic learning settings. For instance, they are used both with regular time-series data as well as with image-based AI methods to obtain accurate dynamical models directly from video. This aids in the spatial awareness of AI which is essential in a multitude of robotic applications and fluid systems as seen in the video below.
The system simulation on the left in the video takes multiple days to complete whereas the AI-based model on the right (CNN encoder) takes no more than a second to complete. This speed shows that AI can understand these fluid systems in a highly efficient manner. This effectiveness enables many applications: from improving system analysis and design to real-time control.
Fast adoption
Beintema thus has shown that with the thoughtful application of AI, many highly complex processes can be efficiently understood. Lastly, his developed methods are publicly available in an off-the-shelf software toolbox deepSI to allow fast adoption by the industry.
Title of PhD thesis: Data–driven Learning of Nonlinear Dynamic Systems: A Deep Neural State–Space Approach. Supervisors: Prof. Roland Toth, and Dr. Maarten Schoukens.