Deep learning for robust and efficient design in cyber-physical systems
Xian Yeow Lee · 2022
Recent developments in machine learning, particularly deep learning, have accelerated due to the combined innovations in computational throughput and algorithmic advancements. Simultaneously, many traditional systems (i.e., engineering systems, manufacturing systems) are also transforming into cyber-physical systems (CPS), with hardware and software components operating in tandem due to the declining cost of digitization. This transformation has created an abundance of opportunities to adapt deep learning approaches to optimize various aspects of these systems by leveraging the data these CPS generates. While deep learning has found successful adoption in areas such as computer vision and natural language processing, the potential benefits of applying deep learning to engineering design in CPS are still mostly untapped due to several existing challenges. Some of the challenges that impede the widespread adoption of deep learning for engineering design in CPS are (i) the existence of combinatorially-large design space, (ii) the ability to incorporate design constraints and domain-specific knowledge, (iii) the availability and cost of high-quality data acquisition, iv) identification of possible failure modes and v) robustifying the designs. While these challenges exist in conventional engineering design, methods to resolve or circumvent these challenges have been well developed based on years of experience. Addressing these challenges for deep learning used for engineering design in CPS will enable the development of deep learning-based data-driven methods to further optimize and support the engineering design processes. This dissertation includes multiple case studies of engineering design in CPS and develops various deep learning-based frameworks that address the challenges discussed above while simultaneously enhancing the design process. The first study involves the development of a supervised deep learning model that is applied to a niche additive manufacturing system for automated parameter selection and quality control. The next case study explores the use of deep reinforcement learning for a sequential design problem in a microfluidic device design example. Following that, a generative modeling framework for designing material microstructures will be presented. In the second part of the dissertation, threat models were developed for deep reinforcement learning-based controllers to identify potential failure modes in the controller designs for several robotic simulations. The final case study presents a possible approach for robustifying reinforcement learning-based controllers with application to power systems. Collectively, these case studies demonstrate the various ways in which deep learning can be adopted to circumvent existing challenges while improving different aspects of the engineering design problem.