適用於生理訊號監護之抗雜訊壓縮感知還原晶片設計

Noise-Tolerant Compressive Sensing Reconstruction Engine Designs for Real-Time Physiological Signals Monitoring

陳庭笙(Ting-Sheng Chen)

指導教授 ： 吳安宇

國立臺灣大學/電機資訊學院/電子工程學研究所/博士(2018年)

https://doi.org/10.6342/NTU201804128

摘要

隨著高齡化社會以及慢性病增長的趨勢下，居家照護的需求大增。其中如何整合可攜式感測器與無線通訊來實現遠程生理訊號監護系統(physiological signals telemonitoring system)，正是實現居家照護的重點方向。目前人體端可攜式無線照護系統面臨到有限的電池容量以及頻寬限制等技術問題，透過低複雜度生理訊號壓縮技術，可以同時解決以上兩個問題，因此開發一個有效的生理訊號壓縮與即時監控系統為本論文主要研究重點。
在遠程生理訊號監護系統的訊號處理方面，系統須連續不斷的偵測生理訊號以提供即時的病情監控，而眾多的訊號將大量消耗系統的頻寬以及功耗。面對此問題，我們提出利用最近期的訊號處理技術-壓縮感知(compressive sensing)來解決，壓縮感知技術是2004年由Donoho教授和Candès教授等人在泛函數分析的基礎上，結合了訊號稀疏特性，所提出的突破性訊號取樣理論。壓縮感知可以將高維度的稀疏訊號，透過測量矩陣取得低維度的測量值，因此系統只需要以低維度的訊號做傳遞，等需要時再利用壓縮感知重建演算法將低維度的取樣重建回高維的訊號。利用壓縮感知技術，我們能提供遠程生理訊號監護系統更有效率的訊息傳遞方式。壓縮感知是新興的訊號處理技術，吸引了許多研究者的目光，它能以遠低於傳統奈奎斯理論的取樣量來取樣訊號，並將訊號由這些較少的取樣中還原。本篇論文將著重於訊號還原的演算法。
然而，在將現有技術應用到後端重建訊號過程，遇到了三個挑戰：1) 重建演算法複雜度高，若需達到多通道、多病人多訊號即時重建，現有軟體的吞吐量與能源效率不佳；2) 量測雜訊是不可避免的問題。少量的量測雜訊，就能造成現有硬體的還原引擎還原失敗；3) 現有還原引擎面臨彈性與吞吐量的取捨問題，若要能支援大範圍且動態可調的參數，則無法兼顧高吞吐量與能源效率。
因此，本論文提出了三項主題來解決上述問題：第一部分是抗雜訊還原演算法，首先，我們發現稀疏值是抗雜訊的關鍵，因此提出了一個基於殘差之稀疏值估計技術(residual-based sparsity estimation technique)，能在未知雜訊等級時又能估計出有效的稀疏值，跟一個改良的半盲解演算法，稱為考量稀疏值之子空間追蹤演算法(sparsity-aware subspace pursuit，SASP)。接著，我們結合基於殘差之稀疏值估計技術與考量稀疏值之子空間追蹤演算法，提出一個嶄新的壓縮感知還原演算法，稱為稀疏值估計匹配追蹤演算法(sparsity estimation matching pursuit，SEMP)。此演算法有幾個特點。其一，它是個盲解演算法，不需要額外的稀疏值輸入。其二，即使不需要額外的稀疏值，此演算法在抗雜訊的表現上，實驗結果比現有演算法還要良好。其三，此演算法擁有低的運算複雜度。第二部分是硬體架構設計，在實作抗雜訊演算法的架構上，我們利用隨機梯度架構來達到高吞吐量與高彈性參數設計，並將兩個演算法設計在同一個硬體上，達成高面積效率。此外，三個優化技術得以提升6.3倍的吞吐量與能源效率。第三部分是晶片實作，我們將提出的還原演算法實作於晶片設計中，晶片量測結果得知，我們提出的還原晶片可以提供每秒232-1996 千取樣的還原吞吐量、消耗不到30mW、8dB以上優異抗躁性、支援大範圍且完全可調參數之設計。
總結來說，有了本論文所提出的盲解低複雜度抗雜訊壓縮感知還原演算法及晶片系統實作，可以有效地解決壓縮感知技術應用到實際的生理訊號監護系統時所遇到的問題，使壓縮感知技術更適合於實際遠距醫療應用。

關鍵字

壓縮感知 ； 電路與系統

並列摘要

With the aging of society and rising of chronic diseases, the demand for home care has been increasing substantially. How to integrate the portable sensors and wireless communication to implement the physiological signals telemonitoring system is the key to home care. Currently, the portable physiological signals telemonitoring system in the body side is facing the problems of limited battery life and limited bandwidth. Physiological signals compression techniques are employed to address these two issues. As a consequence, our goal is to develop an effective physiological signals compression system.
In the signal processing aspect of telemonitoring system, the system needs to detect various physiological signals continuously and provides real-time condition monitoring. These numerous signals consume large bandwidth and power in the system. Faced with such dilemma, we propose to exploit recent signal processing technique, compressive sensing, to resolve these problems. Compressive sensing (CS) is the groundbreaking signal sampling theorem proposed by professor Donoho and Cades based on the functional analysis and sparse signal in 2004. Compressive sensing obtains low dimensional measurements by sampling on high dimensional sparse signals with measurement matrix. Thus, the system only transmits low dimensional signals, while the original high dimensional signals can be recovered by CS reconstruction algorithm. With compressive sensing, we can provide a more efficient way to transmit information in physiological signals telemonitoring system. This thesis focus on the CS reconstruction algorithm.
When CS is applied to the physiological signals compression, the encountering challenges of CS reconstruction include: 1) The computational complexity of CS reconstruction algorithms is too high, as a result, hardware implementation is necessary for multiple channel/patient signals monitoring; 2) Inevitable measurement noise destroy the signal sparsity thus degrading the reconstruction quality, and 3) Existing CS reconstruction engine cannot support robust reconstruction, adjustable parameters, and high throughput/energy efficiency.
There are three main topics in this work. First, we find that the sparsity is the key factor to cope with the measurement noise. As a result, we propose a residual-based sparsity estimation technique that can estimate the sparsity order accurately under noisy scenario. We also propose an improved non-blind algorithm called sparsity-aware subspace pursuit (SASP) algorithm. Then, we propose a novel sparsity estimation matching pursuit (SEMP) algorithm that combining the residual-based sparsity estimation technique and SASP algorithm. There are some features in the proposed SEMP algorithm. First, it is a blind algorithm which is no need of explicit sparsity. Second, although without input of sparsity, its performance in noise-resilience is even better than the state-of- the-art non-blind algorithm. Last but not least, the SEMP has the properties of small overhead and low-complexity.
In the second part of this thesis, we design the hardware architecture for this noise-tolerant CS reconstruction algorithm. We use stochastic gradient descent method to achieve high throughput rate as well as large/flexible parameter design. Then, we implement two reconstruction algorithm into the folded architecture thus achieving area efficiency. In the third part of this thesis, we implement this CS reconstruction algorithm with 40nm CMOS technology. The measurement results show that it can achieve 232-to-1996KS/s, less than 30mW, more than 8dB SNR gain, large-scale and fully reconfigurable parameter setting.
In summary, the 8.66mm2 CS reconstruction engine can provide timely physiological signal reconstruction for data collected from CS-based wireless biosensors under noisy conditions, making intelligent patient monitoring a reality.

並列關鍵字

Compressive Sensing ； Circuits and Systems

參考文獻

[1] J. M. Smith, “Wireless Health Care: Wireless technologies are about to transform health care, and not a moment too soon,” IEEE Spectrum, Sept. 2011.

[2] http://www.spectrum.ieee.org/biomedical/devices/wireless-health-care/

[3] World Health Organization (WHO), “Telemodicine: Opportunities and developments in memer states,” Gobal Observatory for eHealth series, vol. 2, pp. 1-96, Jan. 2011. http://www.who.int/goe/publications/en/

[4] F. Touati, and R. Tabish, “U-Healthcare System: State-of-the-Art Review and Challenges,” Journal of Medical Systems, pp.1–20, May 2013.

[5] W. Cheng, M.F. Yeha, K. C. Changa, R. G. Leeb, “Real-time ECG telemonitoring system design with mobile phone platform”, Measurement, pp. 463-470, May. 2008