Software-Defined Control (SDC) is a revolutionary methodology for controlling manufacturing systems that uses a global view of the entire manufacturing system, including all of the physical components (machines, robots, and parts to be processed) as well as the cyber components (logic controllers, RFID readers, and networks). As manufacturing systems become more complex and more connected, they become more susceptible to small faults that could cascade into major failures or even cyber-attacks that enter the plant, such as, through the internet. In this project, models of both the cyber and physical components will be used to predict the expected behavior of the manufacturing system. Since the components of the manufacturing system are tightly coupled in both time and space, such a temporal-physical coupling, together with high-fidelity models of the system, allows any fault or attack that changes the behavior of the system to be detected and classified. Once detected and identified, the system will compute new routes for the physical parts through the plant, thus avoiding the affected locations. These new routes will be directly downloaded to the low-level controllers that communicate with the machines and robots, and will keep production operating (albeit at a reduced level), even in the face of an otherwise catastrophic fault. These algorithms will be inspired by the successful approach of Software-Defined Networking. Anomaly detection methods will be developed that can ascertain the difference between the expected (modeled) behavior of the system and the observed behavior (from sensors). Anomalies will be detected both at short time-scales, using high-fidelity models, and longer time-scales, using machine learning and statistical-based methods. The detection and classification of anomalies, whether they be random faults or cyber-attacks, will represent a significant contribution, and enable the re-programming of the control systems (through re-routing the parts) to continue production. The manufacturing industry represents a significant fraction of the US GDP, and each manufacturing plant represents a large capital investment. The ability to keep these plants running in the face of inevitable faults and even malicious attacks can improve productivity -- keeping costs low for both manufacturers and consumers. Importantly, these same algorithms can be used to redefine the production routes (and machine programs) when a new part is introduced, or the desired production volume is changed, to maximize profitability for the manufacturing operation.

Recent years have seen an explosion in the use of cellular and wifi networks to deploy fleets of semi-autonomous physical systems, including unmanned aerial vehicles (UAVs), self-driving vehicles, and weather stations to perform tasks such as package delivery, crop harvesting, and weather prediction. The use of cellular and wifi networks has dramatically decreased the cost, energy, and maintenance associated with these forms of embedded technology, but has also added new challenges in the form of delay, packet drops, and loss of signal. Because of these new challenges, and because of our limited understanding of how unreliable communication affects performance, the current protocols for regulating physical systems over wireless networks are slow, inefficient, and potentially unstable. In this project we develop a new computational framework for designing provably fast, efficient and safe protocols for the control of fleets of semi-autonomous physical systems. The systems considered in this project are dynamic, defined by coupled ordinary differential equations, and connected by feedback to a controller, with a feedback interconnection which has multiple static delays, multiple time-varying delays, or is sampled at discrete times. For these systems, we would like to design optimal and robust feedback controllers assuming a limited number of sensor measurements are available. Specifically, we seek to design a class of algorithms which are computationally efficient, which scale to large numbers of subsystems, and which, given models of the dynamics, communication links, and uncertainty, will return a controller which is provably stable, robust to model uncertainty, and provably optimal in the relevant metric of performance. To accomplish this task, we leverage a new duality result which allows the problem of controller synthesis for infinite-dimensional systems to be convexified. This result allows the problem of optimal and robust dynamic output-feedback controller synthesis to be reformulated as feasibility of a set of convex linear operator inequalities. We then use semidefinite programming to parametrize the set of feasible operators and thereby test feasibility of the inequalities with little to no conservatism. In a similar manner, estimator design and optimal controller synthesis are recast as semidefinite programming problems and used to solve the problems of sampled-data and systems with input delay. The algorithms will be scalable to at least 20 states and the controllers will be field-tested on a fleet of wheeled robotic vehicles.