Data-Driven Cyber-Physical Systems via Real-Time Stream Analytics and Machine Learning

Emerging distributed cyber-physical systems (CPSs) integrate a wide range of heterogeneous components that need to be orchestrated in a dynamic environment. While model-based techniques are commonly used in CPS design, they be- come inadequate in capturing the complexity as systems become larger and extremely dynamic. The adaptive nature of the systems makes data-driven approaches highly desirable, if not necessary.Traditionally, data-driven systems utilize large volumes of static data sets to extract models and predictions of physical processes. However, in emerging CPS, networked sensors provide continually streaming data, creating an essentially infinite source of information. Processing data in batches is no longer a viable option: streams are most valuable when processed on-line, allowing actionable information to be gathered just as the data becomes available. This fundamental shift from big data to infinite data, while having great potential to enable smarter systems, also poses unique challenges. Computation models that capture the integration of streaming data into CPS design become a key requirement for systems to learn, adapt, and evolve in real-time.This thesis explores methodologies for developing data-driven CPSs that integrate model-based design and real-time stream analytics in a modular way. The key modeling framework to be introduced is the aspect-oriented modeling (AOM) paradigm, which leverages the principle of separation-of-concerns in actor-oriented de- sign. Aspects are useful for representing cross-cutting concerns in complex system architectures, as first introduced by the aspect-oriented programming paradigm in object-oriented design. AOM applies this idea to actor-oriented design, creating aspects that enable representation of modular concerns in a complex system model. In data-driven CPS, the introduced aspects can be leveraged to process streaming data, extract actionable information, and incorporate these into the system workflow in a way that preserves model semantics and modularity. To address information extraction from streaming data, we propose the use of aspects that implement Dynamic Bayesian Network based algorithms for machine learning and optimization. Specifically, we introduce an actor-oriented toolkit that enables dynamics and sensing models to be composed with inference, Bayesian learning, and optimization algorithms, and present comprehensive case studies on cooperative mobile robot control. We additionally study the use of streaming data for control of dynamic networked CPS in the context of home automation, and present an overview of the use cases of aspects in actor-oriented CPS development