Model-Based Design Optimization and Simulation Techniques for Dynamic, Data-Driven Application Systems
Bhattacharyya, Shuvra S
MetadataShow full item record
Dynamic Data Driven Application Systems (DDDAS) are an important class of systems in which computations on data and instrumentation components for acquiring the data are incorporated within a feedback control loop. In DDDAS, the modeling and data-driven adaptation of instrumentation is incorporated as an important aspect of the design process. Due to its potential to enhance capabilities of accurate analysis, dynamic decision making, and scalable simulation, the DDDAS paradigm plays an increasingly important role in innovative systems for a wide variety of applications. This thesis develops new model-based, software design tools to support the design and implementation of DDDAS. The methods are developed in the context of two application domains in which DDDAS principles are highly relevant --- multispectral/hyperspectral video processing, and wireless-integrated factory automation systems. Recent advances in multispectral and hyperspectral video capture technology along with system design trade-offs introduced by these advances present new challenges and opportunities in the area of DDDAS for video analytics. Video analytics plays an important role in a wide variety of defense-, monitoring- and surveillance-related systems for air and ground environments. In this context, multispectral video processing is attracting increased interest in recent years, due in part to technological advances in video capture. Compared with monochromatic video, multispectral video offers better spectral resolution, and different bands of multispectral video streams can enhance video analytics capabilities in different ways. Video processing systems that incorporate multispectral technology involve novel trade-offs among system design complexities such as spectral resolution, equipment cost, and computational efficiency. The design space of multispectral video processing systems is enriched by considering only the required subset of spectral bands to process as a parameter that can be adjusted dynamically based on data characteristics and constraints involving accuracy, communication, and computation. Based on this view of selectively-processed bands from multispectral video data, we introduce in this thesis a novel system design framework for dynamic, data-driven video processing using lightweight dataflow (LD) techniques. Our proposed framework, called LDspectral, applies LD, which is an approach for model-based design of signal and information processing systems. LD facilitates efficient and reliable real-time implementation. LD is ``lightweight'' in the sense that it is based on a compact set of application programming interfaces, and can be integrated relatively easily into existing design processes. We develop a framework for adaptively configuring multispectral video processing configurations in LDspectral, and develop a prototype implementation using LD methods that are integrated with OpenCV, which is a popular library of computer vision modules. We demonstrate and evaluate the performance of LDspectral capabilities using a background subtraction application. As compared to a standard video processing pipeline, the capabilities in LDspectral for optimized selection and fusion of spectral bands enhance trade-offs that can be realized between video processing accuracy and computational efficiency. Using the DDDAS paradigm, the elements of sensor measurements, statistical processing, target modeling, and system software are analyzed by frequency bands, video analytics, environmental analysis, and dataflow techniques, respectively. In this thesis, the LDspectral framework is also extended to hyperspectral video, which offers great spectral resolution and has significant potential to enhance the effectiveness of information extraction from image scenes. An important challenge in the development of hyperspectral video systems is managing the high computational load and storage requirements required to process the large volumes of data that are acquired by these systems. We also investigate DDDAS-inspired methods in context of distributed, smart factory systems that are equipped with wireless communication capability. We refer to this class of systems as wireless-integrated factory systems (WIFSs). An important challenge in the development of this class of systems is ensuring reliable, low latency communication under the harsh wireless channel conditions of factory environments. To support the application of the DDDAS paradigm in WIFSs, we develop a model-based software tool for design space exploration. We refer to this tool as the Wireless-Integrated factory System Evaluator (WISE). WISE supports the rapid simulation-based evaluation of interactions among the placement of factory subsystems, the partitioning of factory subsystems into nodes of a wireless network, the performance of the wireless network, and overall factory system performance. WISE also incorporates a new graphical model called the cyber-physical flow graph, which provides integrated modeling for the flow of physical entities (such as parts that are processed in a factory) and the flow of information. The cyber-physical flow graph also models distributed flows in which information is communicated across multiple network nodes.