Managing heterogeneous resources for dynamic energy-efficient sensing

energy-efficient sensing;sensor utilization management;embedded sensors;acoustic monitoring

Jun, David

关键词: energy-efficient sensing;    sensor utilization management;    embedded sensors;    acoustic monitoring;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/49694/David_Jun.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Next-generation embedded devices are expected to pervasively extract information from the world around us. In particular, growing interest in mobile devices beyond the smart phone imply that the need for context awareness and information extraction is more important than ever before.This thesis considers the optimization in the utilization and design of sensing resources on a single device, in order to extract high-quality information at energy consumption rates that are dramatically lower than what is expected today. The key insight is to match device resources to the information available in the environment, which requires active resource management along with a diverse set of efficient hardware and software components.The problem of matching device resources to the information available in dynamic environments is fundamentally difficult because the amount of effort and energy spent on sensing is typically related to the quality of the data acquired, which determines the accuracy of understanding and predicting the state of the environment. This thesis develops the analysis required to understand optimal trade-offs between performance and energy consumption for a system with a given set of sensing and processing resources. The utilization problem is mapped to a partially observable Markov decision process (POMDP) and the appropriate mapping is derived in order to leverage state-of-the-art POMDP numerical solvers to generate optimal resource-scheduling policies.Developing a tool to determine the optimal achievable performance/energy trade-off for a given set of resources enables system designers to understand the inefficiencies of heuristic sensing strategies, and also to propose more aggressive trade-offs for new sensors, signal chains, and algorithms. We present a case study of our approach to system design, using an acoustic wildlife monitoring task as the application driver. A hidden Markov model (HMM) approach to pattern recognition is adopted, where event arrivals and bird songs are modeled using a 30-state Markov chain, with transition parameters learned from data. In order to characterize the signal fidelity and energy costs associated with sensing and processing resources, the CheetahCub testbed is developed, which combines a TI MSP430 low-power microcontroller with dynamic voltage scaling driven by algorithmic processing demands. The testbed achieves 16x energy scalability, ranging from simple energy calculations to more sophisticated signal processing. We also discuss the EFM testbed, based on the 32-bit ARM Cortex-M3 processor, which proves to be 2.5x more efficient than the CheetahCub testbed when high processing capabilities are required.In both testbeds, the data acquisition front-end is shown to be the energy bottleneck. Thus, a novel signal chain for acoustic sensing is proposed. The proposed signal chain replaces the typical preamplifier circuit and ADC with a digital noise floor tracker and analog comparator. The sensing package consumes 10x less energy than a traditional microphone circuit but at the expense of degraded signal fidelity.Observation models are learned from data for the sensing actions developed in this thesis. The procedure for mapping the problem to a POMDP to generate optimal scheduling policies is demonstrated, and our approach to system design is validated by evaluating the optimal performance/energy trade-off achieved by our system. Including the proposed nano-power acoustic sensor demonstrates an order of magnitude reduction in total system energy consumption, relative to an efficient approach based on cascading signal models for energy-aware detection. This process demonstrates the powerful synergy achieved by utilizing theory to enable systematic evaluation of aggressive, innovative sensor-design trade-offs. Our optimal scheduling framework is also used to study the efficiency of an intuitive wakeup mechanism, demonstrating that heuristic design may be missing out on 2x-4x energy savings, compared to optimal scheduling.In closing, this thesis challenges conventional wisdom that devices must be designed to sleep in ultra low-power sensing applications. We demonstrate that this does not necessarily have to be true, if one can combine innovative sensing hardware with clever resource management to match the information available in dynamic environments. This will become increasingly relevant as emerging applications in wearable computing and the Internet of Things demand high-quality information and context awareness.