From Model-Based Foundations to Data-Driven Intelligence: Toward Resilient Networked Cyber-Physical Systems

ABSTRACT

Networked Cyber-Physical Systems (N-CPS) consist of intelligent computing/control (“cyber”) elements and real-world (“physical”) subsystems, interconnected (“networked”) through various communication and physical processes. These systems are pervasive in our environment, playing integral roles in critical infrastructure, such as multi-robot systems, vehicular platoons, supply chain networks, smart grids, and epidemic spreading networks. Therefore, careful design of such N-CPS is paramount to promote their overall stability, robustness, safety, and resilience.

To address these requirements, our research has developed an innovative model-based framework for co-designing controllers and network topology in N-CPS. This co-design framework is not only computationally efficient but also operationally compositional, enabling it to handle large-scale N-CPS reconfigurations efficiently and thereby enhance resilience. Furthermore, it not only preserves N-CPS stability but also optimizes the impact of disturbances on performance. Moreover, it provides insightful control and topology designs and is broadly applicable across a wide range of N-CPS.

Complementing the developed model-based framework, we have also developed data-driven AI techniques for N-CPS, utilizing distributed Graph Recurrent Neural Networks (GRNNs) to learn optimal controllers online, exploiting only locally available data and neighbor exchanges. This method is fully decentralized and provides control-theoretic stability and robustness guarantees, exploiting sector-bounded and slope-restricted characteristics of GRNN activation functions. We further address the post-fault/attack recovery of N-CPS by learning the unknown/altered dynamics online with Recurrent Equilibrium Networks (RENs) and by embedding a funnel-based recovery controller that drives the state back to an invariant tube around the nominal trajectory, thereby guaranteeing safety and enhancing resilience.

In this talk, I will start by introducing the dissipativity principles that underpin the model-based co-design framework and illustrate their application to representative N-CPS scenarios. I will then demonstrate how these ideas are extended to the data-driven setting through GRNN-based distributed optimal control and REN-based online recovery, enabling stable, robust, safe, and resilient operation.

BIOGRAPHY

Shirantha Welikala is currently an Assistant Professor in the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ, USA (joined Fall 2023). He received the B.Sc. degree in Electrical and Electronic Engineering from the University of Peradeniya, Peradeniya, Sri Lanka, in 2015 and the M.Sc. and the Ph.D. degrees in Systems Engineering from Boston University, Brookline, MA, USA, in 2019 and 2021, respectively. From 2015 to 2017, he was a Temporary Instructor/Research Assistant in the Department of Electrical and Electronic Engineering at the University of Peradeniya, Sri Lanka. From 2021 to 2023, he was a Postdoctoral Research Fellow in the Department of Electrical Engineering, University of Notre Dame, Notre Dame, IN, USA. His main research interests include control and optimization of cooperative multi-agent systems, control of networked systems, passivity-based control, control and topology co-design, machine learning, robotics, and smart grid. He has received several awards, including the 2015 Ceylon Electricity Board Gold Medal, the 2019 and 2023 President's Awards for Scientific Research in Sri Lanka, the 2021 Outstanding Ph.D. Dissertation Award in Systems Engineering at Boston University, the 2022 Best Paper Award at the 30th Mediterranean Conference on Control and Automation, and the Best Paper Finalist Award at the American Control Conference, 2025.