Self Learning Control: a breakthrough in efficient, low-emission automotive powertrains
Prasoon Garg defended his PhD thesis at the Department of Mechanical Engineering on November 6th.
In partnership with DENSO Aachen, for his PhD research Prasoon Garg explored Self Learning Control methods for automotive powertrains, using Machine Learning to automate and optimize control calibration. As road transport is a major source of greenhouse gas emissions, the automotive industry faces the critical challenge of developing cleaner, more efficient powertrain systems. Through innovative techniques like Long Short-Term Memory (LSTM) neural networks and Reinforcement Learning, this work achieves faster, safer calibration processes, including real-world validation on a vehicle test bench. With its patented auto-calibration strategy, Garg demonstrates the potential of SLC to enhance efficiency, reduce emissions, and cut development time鈥攎arking a significant step forward in AI-driven automotive innovation.
The last few decades have seen an increase in the complexity of the automotive powertrains and this trend is expected to continue in the future. New technologies include advancements in existing internal combustion engines, advanced combustion technologies and increasing levels of electrification. This will go hand in hand with an increase in the control development time for these systems. Therefore, there is an increasing need for new control calibration approaches that can minimize the calibration effort and achieve robust performance in the real-world for future powertrains.
Optimal control settings
In close collaboration with the Energy Systems R&D team of our German/Japanese industry partner DENSO Aachen Engineering Center, this research of Prasoon Garg focused on developing Self Learning Control methods, which show potential to minimize the calibration effort and improve robust performance in real-world operations. In Self Learning Control, the control method can learn optimal control settings by autonomously interacting with the system, which can significantly reduce the calibration effort by minimizing the expert effort to a great extent. In pursuit of developing new Self Learning Control methods, this work investigated Machine Learning methods, which comprise a variety of algorithms for solving complex problems, including regression, classification, clustering and decision-making.
Transient performance is key
Powertrain systems such as diesel engines, are characterized by their highly dynamic operation, so transient performance is key. Currently applied control methods are unsystematic and require tuning of an increasing number of look-up tables and of parameters in the applied models. The transient performance of the system has been improved by developing preview control method using Long Short-Term Memory (LSTM) neural networks. The challenging parameter tuning problem has been addressed by systematically selecting parameters that optimize not only the LSTM neural network model accuracy but also control system performance, calibration effort and computational requirements.
Reinforcement Learning
Further reduction in the expert effort is made by developing Reinforcement Learning-based methods for the automation of the calibration process and learning optimal control settings on a real vehicle. A novel exploration method has been developed in this work for safe and time-efficient auto-calibrated, which is experimentally validated on a real vehicle test-bench. A patent application has been filed in German Patent and Trade Mark Office. These experimental demonstrations place this work among the first few attempts to apply Reinforcement Learning on a physical system with safety constraints.
Title of PhD thesis: . Supervisors: Prof. Frank Willems, and Assistant Prof. Emilia Silvas.
PhD in the picture
What was the most significant finding from your research, and what aspects turned out to be most important to you?
That would be the Self Learning control method using Reinforcement Learning. In this topic, I addressed a key challenge of realizing safe learning using Reinforcement Learning for an engineering application. Moreover, I have developed an automated calibration method to learn directly from the physical system without a prior system model. For this, I have developed a novel safe and time-efficient exploration method.
The most important aspect to me in this topic is that I made a contribution to both academic and industry with the novel exploration method and automated calibration approach, respectively. Another aspect important to me, is that I made an experimental demonstration of the developed method on a vehicle test bench at DENSO, which provided me insights into practical aspects. Lastly, along with DENSO, we have made a patent application for this topic in German Patent and Trade Mark Office.
What was your motivation to work on this research project?
The opportunity to learn and address challenging problems from the worlds of academic research and automotive industry (application) as this project was a close collaboration between 果冻传媒 and DENSO Aachen Engineering Center in Germany.
What was the greatest obstacle that you met on the PhD journey?
At times bringing different stakeholders with varying expectations and requirements on a common ground. I believe this is quite natural when the worlds of research and application come together.
What did you learn about yourself during your PhD research journey? Did you develop additional new skills over the course of the PhD research?
That I can be very persistent towards my target and this helps me in working out the best possible solution irrespective of the situation.
During the course of my PhD, I gained multiple skills owing to collaboration between 果冻传媒 and DENSO. It includes handling stakeholders with varying technical backgrounds, which requires both zooming in and out on levels of detail, proactive project planning and hands-on experience with implementing and testing Machine Learning-based control methods on the vehicle rapid prototyping control systems.
What are your plans for after your PhD research?
Since last 2 months, I am working at DAF Trucks, Eindhoven as a Function Design Engineer in Global Engines department for developing control concepts.