Our paper on ``Federated learning of predictive models from federated Electronic Health Records''...
SENSORS: A Control and Optimization Science Base for Sensor Networks in Adverse and Stochastic Environments
Funding Agency: National Science Foundation, Directorate for Engineering, Division of Design and Manufacturing Innovation, Sensors and Sensor Networks Initiative Project.
Award Number: DMI-0330171.
Principal Investigators: C. G. Cassandras, J. Baillieul, D. Castanon, and Yannis Paschalidis at Boston University (with R. Gao, W. Gong, and A. Deshmukh at UMass Amherst).
The goal of this project is to contribute to the foundations of a control and optimization science base for Sensor Networks (SNETs) viewed as complex systems operating in a stochastic and potentially adverse environment. The project is intended to address fundamental research issues, while maintaining a focus on a specific target application domain – a manufacturing enterprise – to capitalize on PI expertise and generate concrete, implementable results. It is the premise of this proposal that in a SNET: (i) sensors must be able to operate with limited power/bandwidth; (ii) sensors must possess some embedded intelligence allowing them to alter their data collection/transmission operation based on the perceived state of the environment; (iii) data from distributed groups of sensors, operating asynchronously, must be integrated and processed in ways that compensate for potentially poor quality and delayed or partial information; and (iv) processed data must be able to drive effective control actions, while also avoiding false alarms.
The proposed research plan adopts three complementary, interconnected views: (i) a SNET as a collection of elements that need to communicate among them; (ii) a SNET as a dynamic system whose control depends on a thorough understanding of interacting state dynamics; and (iii) a SNET as a multi-time scale distributed system that requires a layer of coordination. Under (i), the project will address problems of allocating communication resources and managing message exchange in a highly distributed and asynchronous setting while taking into account energy and bandwidth availability. Under (ii), a hybrid system modeling framework will be developed, capturing the interaction of time-driven physical dynamics and event-driven operational dynamics of sensors, and, hence, associated performance tradeoffs leading to dynamic optimization problems. Under (iii), the problems addressed are the registration and fusion of disparate sensor data, and the development of appropriate models for hierarchical multi-time scale systems that rely on these data. The solution of these problems entails considerable analytical and computational challenges that will be faced through novel modeling, estimation, and optimization approaches. A key component of the project is the deployment of an experimental testbed. It will be partly deployed at the University of Massachusetts (UMass) on a manufacturing machine platform using mechanical sensors and low-power digital signal processors. It will leverage the expertise of the PIs who have previously demonstrated millimeter-sized sensors inserted into ball bearings for condition monitoring. Another SNET will be formed at Boston University (BU) using an existing set of mobile robots as multi-sensor platforms wirelessly networked with a group of laptop computers serving as \processing” nodes. A Linux computer grid at UMass will remotely connect the two SNETs for research on network communications. Research results will include new sensors for manufacturing systems, new energy saving protocols for wireless SNETs, and new modeling and control approaches for complex engineering systems, all validated on this testbed.
The intellectual merit of this proposal is threefold. First, it will contribute to the conceptualization and analysis of fundamental modeling and control issues for SNETs that will lead to a thorough understanding of their potential and limitations. Second, it will bring novel ideas and methods to the overall effort of developing and deploying SNETs (e.g., the theory of discrete event and hybrid systems, concurrent estimation, distributed and asynchronous optimization methods); the use of stochastic optimization in particular is viewed as an essential ingredient for the viability of a system operating under strict energy and bandwidth constraints. Finally, this research will explore two significant shifts in systems engineering with intriguing implications: from strictly time-driven to a more versatile event-driven data sampling and processing approach; and from sensor-poor to data-rich control system design.
Broader impact, dissemination and educational plans. The project will advance the state-of-the-art in all application domains that can benefit from SNETs, primarily focusing on manufacturing where advances will result in increased energy efficiency, accelerated productivity growth, improved product quality, and enhanced workplace safety and security. Other societal applications include transportation safety and homeland security (i.e., proposed work on active surveillance networks). On the educational front, plans include multidisciplinary training of graduate and undergraduate students, creating software modules for educational case studies, and working through the NSF science ambassador award to BU to reach out to high school students. In addition to the usual means of disseminating the outcomes of the proposed work (publications, conferences, etc.), the project will leverage industrial contacts through the BU Center for Information and Systems Engineering and a recent NSF IGERT award to sponsor doctoral student internships and post doctoral fellows. The proposed SNET testbed will be available to the scientific community for collaborative experimentation. Finally, dissemination plans include the organization of an academic workshop on SNETs and an industry-oriented workshop (as part of a regular such series at BU).