CAREER: Building Energy-Efficient IoT Frameworks - A Data-Driven and Hardware-Friendly Approach Tailored for Wearable Applications

职业：构建节能的物联网框架 - 专为可穿戴应用量身定制的数据驱动且硬件友好的方法

• 批准号: 1652038
• 负责人: Fengbo Ren
• 金额: $ 53.42万
• 依托单位: Arizona State University
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2017
• 资助国家: 美国
• 起止时间: 2017-02-15 至 2023-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1652038&HistoricalAwards=false
• 关键词: CAREER; Building; Energy; Efficient; IoT

Sensor energy efficiency is the top critical concern that hinders long-term monitoring in energy-constrained Internet-of-things (IoT) applications. Conventional compressive sensing techniques fail to achieve satisfactory performance in IoT and especially wearable applications due to the lack of prior knowledge about signal models and the overlook of individual variability. The research goal of this CAREER plan is to develop a data-driven and hardware-friendly IoT framework to fundamentally address the unmet energy efficiency need of IoT and especially wearable applications. This will be accomplished by a systematic approach that seamlessly integrates compressive sensing and data analytics in compressed domains using deep learning methods. The proposed research will provide a transformative IoT framework that significantly reduces the data size for transmission from sensors to cloud while improving the overall quality of information delivery and bringing signal intelligence closer to users. The research outcomes will directly impact a variety of IoT applications, such as long-term environmental sensing for monitoring the airborne quality, radiation, water quality, hazardous chemicals, and many other environment indicators, by allowing compressive sensors to be deployed in energy-constrained environments to perform precise information acquisition over a significantly increased time span impossible with existing technologies. The proposed framework will also advance wearable technologies to enable important progress in transforming the existing healthcare model from episodic examination for disease diagnosis and treatment to continuous monitoring for disease prediction and prevention. This will make our healthcare systems more effective and economic and improve the overall quality of living for billions of individuals. The PI will take advantage of his affiliation with the I/UCRC Center for Embedded Systems at ASU to engage industry sponsors to accelerate technology adoption and transfer to benefit the society at large. The PI also plans to undertake an ambitious education program to actively engage and impact a diverse population of K-12, undergraduate, and graduate students to take away the PI?s research and create more values for the community in the long term.The specific research objectives are to 1) formulate problems and develop efficient solvers to construct binary near-isometry embedding matrices to enable effective data compression on sensors through compressive sampling; 2) train deep neuron networks to decode information directly from the compressive samples for on-chip data analytics; 3) prototype the proposed framework in wearable hardware and evaluate the system performance over a variety of physiological signals. The research outcomes will allow future IoT devices to precisely sense and transfer the information of interest specified by users in an energy-efficient manner rather than recording imprecise data in raw forms as in existing approaches. The findings from this research will advance the theory development of data-driven compressive sensing by filling the current knowledge gap on how to design near-isometry embedding matrices with binary constraints that are essential for cost-effective hardware mapping. It will also uncover the intrinsic connections between compressive sensing and deep learning by establishing a viable data analytics solution for decoding high-level information directly from compressive samples. On the integration of research and education, the PI will enhance the current curriculum to better prepare students for careers in both industry and academic. The PI will take advantage of the FURI program at ASU to engage undergraduate students in research to foster their interest and motivation to pursue graduate degrees. ASU has one of the largest Hispanic and Native American student populations in the nation. The PI will make strong personal efforts to encourage the recruitment, retention, and advancement of the underrepresented groups. The PI will also collaborate with the Fulton Engineering Education Outreach office to initiate an exciting high school teacher training program, which aims to increase the level of literacy and interest in STEM fields of a large body of high school students through advanced coursework development.

传感器能效是阻碍能源受限物联网（IoT）应用长期监控的首要关键问题。传统的压缩感知技术由于缺乏关于信号模型的先验知识和忽视个体变异性，在物联网特别是可穿戴应用中无法实现令人满意的性能。该CAREER计划的研究目标是开发一个数据驱动和硬件友好的物联网框架，从根本上解决物联网，特别是可穿戴应用未满足的能源效率需求。这将通过一种系统的方法来实现，该方法使用深度学习方法在压缩域中无缝集成压缩感知和数据分析。拟议的研究将提供一个变革性的物联网框架，大大减少从传感器传输到云的数据大小，同时提高信息交付的整体质量，并使信号智能更接近用户。研究成果将直接影响各种物联网应用，例如用于监测空气质量，辐射，水质，危险化学品和许多其他环境指标的长期环境传感，允许在能源受限的环境中部署压缩传感器，以在现有技术不可能实现的显着增加的时间跨度内执行精确的信息采集。拟议的框架还将推动可穿戴技术的发展，使现有的医疗模式从疾病诊断和治疗的偶发性检查转变为疾病预测和预防的持续监测。这将使我们的医疗保健系统更加有效和经济，并提高数十亿人的整体生活质量。PI将利用他与ASU I/UCRC嵌入式系统中心的联系，与行业赞助商合作，加速技术的采用和转让，以造福整个社会。PI还计划开展一项雄心勃勃的教育计划，积极参与和影响K-12，本科生和研究生的多元化人口，以带走PI？具体的研究目标是：1）提出问题并开发高效的求解器，以构建二进制近等距嵌入矩阵，从而通过压缩采样在传感器上实现有效的数据压缩; 2）训练深度神经元网络，直接从压缩样本中解码信息，用于片上数据分析; 3）在可穿戴硬件中对所提出的框架进行原型设计，并评估系统在各种生理信号上的性能。研究成果将使未来的物联网设备能够以节能的方式精确地感知和传输用户指定的感兴趣的信息，而不是像现有方法那样以原始形式记录不精确的数据。这项研究的结果将通过填补目前关于如何设计具有二进制约束的近等距嵌入矩阵的知识空白来推进数据驱动压缩感知的理论发展，这对于具有成本效益的硬件映射至关重要。它还将通过建立一个可行的数据分析解决方案来直接从压缩样本中解码高级信息，从而揭示压缩感知和深度学习之间的内在联系。在研究与教育的整合方面，PI将加强现有的课程，以更好地为学生在工业和学术领域的职业生涯做好准备。PI将利用亚利桑那州立大学的FURI项目，让本科生参与研究，以培养他们攻读研究生学位的兴趣和动力。亚利桑那州立大学拥有全国最大的西班牙裔和美洲原住民学生人口之一。PI将做出强有力的个人努力，以鼓励招聘，保留和代表性不足的群体的进步。PI还将与富尔顿工程教育外联办公室合作，启动一项令人兴奋的高中教师培训计划，旨在通过高级课程开发提高大量高中学生的STEM领域的识字水平和兴趣。

项目成果

Build a Compact Binary Neural Network through Bit-level Sensitivity and Data Pruning

• DOI: 10.1016/j.neucom.2020.02.012
• 发表时间: 2018-02
• 期刊: Neurocomputing
• 影响因子: 6
• 作者: Yixing Li;Fengbo Ren

Learning in the Frequency Domain

• DOI: 10.1109/cvpr42600.2020.00181
• 发表时间: 2020-02
• 期刊: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)
• 作者: Kai Xu;Minghai Qin;Fei Sun;Yuhao Wang;Yen-kuang Chen;Fengbo Ren

LAPRAN: A Scalable Laplacian Pyramid Reconstructive Adversarial Network for Flexible Compressive Sensing Reconstruction

• DOI: 10.1007/978-3-030-01249-6_30
• 发表时间: 2018-07
• 期刊: ArXiv
• 作者: Kai Xu;Zhikang Zhang;Fengbo Ren

Light-Weight RetinaNet for Object Detection on Edge Devices

• DOI: 10.1109/wf-iot48130.2020.9221150
• 发表时间: 2020-06
• 期刊: 2020 IEEE 6th World Forum on Internet of Things (WF-IoT)
• 作者: Yixing Li;A. Dua;Fengbo Ren

OpenICS: Open Image Compressive Sensing Toolbox and Benchmark

• DOI: 10.1016/j.simpa.2021.100081
• 发表时间: 2021-02
• 期刊: Softw. Impacts
• 作者: Jonathan Zhao;Matthew Westerham;Mark Lakatos-Toth;Zhikang Zhang;Avi Moskoff;Fengbo Ren