Please use this identifier to cite or link to this item:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73162
Title: | 低複雜度之壓縮域學習用於即時無線醫療監控 Low-Complexity Compressed Learning for Real-time Wireless Healthcare Monitoring |
Authors: | Ching-Yao Chou 周敬堯 |
Advisor: | 吳安宇 |
Keyword: | 壓縮感知,壓縮學習, Compressed Sensing,Compressed Learning, |
Publication Year : | 2019 |
Degree: | 博士 |
Abstract: | Long-term health-care monitoring, especially outside the hospital setting, offers great potential to substantially improve patient health, quality of life and outcomes. This potential depends on two aspects: 1) the ability to acquire signals that are informative with respect to patient stage; and 2) the ability to make relevant inference from such signals. Since the electrocardiography (ECG) signal recorded from the electrical activity of the heart over a period of time has been utilized for diagnosis for many diseases, the ECG monitoring is recognized as a promising technique to realize telemedicine.
As wireless wearable physiological sensor nodes are known to be resource-limited, it is a crucial problem to reduce the signal acquisition on these sensing system and enhance the energy efficiency of data transmission. Compressed Sensing (CS) is a revolutionary technology combining both sampling and compression through random projection, which enables sub-Nyquist sampling and low-energy data reduction, resulting in life-time extension of the sensor node. Moreover, health monitoring has emphasized the need for edge computing, leveraging the benefits of analyzing real-time data, without the bandwidth costs that come with sending the data offsite (i.e., to the cloud or the data center). Although CS has successfully enhanced the ability to acquire signals, the low-complexity data analytic for the compressed signals should be explored due to resource-constrained edge device. Nevertheless, the prior CS frameworks encounter two main challenges as follow. On the one hand, although the reconstruction can be extremely energy-intensive and time-consuming; in most cases, we only want to analyze the problematic physiological signals. However, the compressed signals are reconstructed before classification in reconstructed learning. Hence, large efforts are wasted in recovering uninterested (or normal) signals which are far more than interested (or high-risk) signals. On the other hand, the availability and transfer of personal information has also posed great concerns for potential privacy leakage. To make the transmitted data useful only for the intended utility but not easily repurposed into privacy intrusion, the concept of Privacy-Preserving Data Mining (PPDM) has recently been proposed. Unfortunately, there is no CS-based PPDM which makes the encrypted compressed signals able to be classified but unable to be reconstructed. To overcome the aforementioned challenges, this dissertation presents compressed learning based on information splitting concept. On the one hand, to eliminate the abundant cost of recovering irrelevant data, a low-complexity two-stage classification-aided reconstruction (TS-CAR) framework is proposed. TS-CAR can not only directly classify the compressed signals, but also accelerate the speed of on-demand reconstruction by using the discriminative information in classification. Furthermore, we design the hardware architecture for TS-CAR algorithm. The proposed intelligent CS reconstruction engine is implemented in TSMC 40 nm CMOS technology. On the other hand, to achieve CS-based PPDM, a low-complexity privacy-preserving compressed analysis (PPCA) framework is proposed. The encryption and decoding protocol in PPCA only retains the discriminative information for classification; therefore, the system can achieve higher privacy without sacrificing the data usage. In summary, this dissertation focuses on low-complexity analyzing CS measurement in edge computing for future intelligent and secured wireless health-care monitoring. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/73162 |
DOI: | 10.6342/NTU201901257 |
Fulltext Rights: | 有償授權 |
Appears in Collections: | 電子工程學研究所 |
Files in This Item:
File | Size | Format | |
---|---|---|---|
ntu-108-1.pdf Restricted Access | 13.98 MB | Adobe PDF |
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.