Ultra-Low Power Feature-to-Digital Converters for IoT Sensor Nodes

Internet-of-things (IoT) devices powered by acoustic processing systems currently have a large share in the biomedical and wearable market. The advance of machine learning provides IoT systems with powerful capability in various applications. Feature extraction is essential for machine learning, and determines the system performance. As many IoT devices have a limited power supply, the power efficiency in computing features dramatically affects the battery lifetime of the device. For portable IoT devices powered by thin-film battery or energy harvester, its power budget is in sub-μW scale for years of operation. This power restriction urges the development of sub-μW acoustic feature extraction circuits.Conventionally, an analog front-end digitizes the raw acoustic signal and then feature extraction is performed in the digital domain. This approach needs the analog-to-digital converter (ADC) and the subsequent digital signal processing (DSP) and communication circuits to operate at the raw signal's rate. Given that acoustic features usually have much lower bandwidth than the raw signal, and the analog circuit exhibits superior efficiency than its digital counterpart in scenarios of relaxed signal-to-noise ratio (SNR) requirements, computing features in the analog domain before digitizing can achieve better power and bandwidth efficiency than the conventional way. A feature-to-digital converter is a circuit that performs analog feature extraction and digitization. However, analog feature extraction often suffers from process variations, which could lead to costly chip-wise calibration or training in a machine learning algorithm. Our goal is to design acoustic feature-to-digital converters under an ultra-low power budget (nW-μW) and to minimize the impact of process variation.Two novel ultra-low power feature-to-digital converters are reported in this thesis. The first one is an Envelope-to-Digital Converter (EDC). The signal's envelope contains important features for a host of applications empowered by machine learning algorithms. The prototype EDC, designed for respiratory monitoring, consumes only 9.6nW. It employs a switched-capacitor (SC) envelope detector to efficiently extract the envelope of a signal as its feature. A ping-pong operation scheme is proposed to integrate the passive envelope detector and the subsequent ADC without needing a buffer, which can otherwise dominate the power budget. In the experiment, the breath rate is successfully measured with envelope signals extracted by EDC from human lung sounds.The second design is a 270nW acoustic feature extractor for always-on voice activity detection (VAD). The design employs a passive SC bandpass filter bank to extract spectral information from noise-corrupted human speech. The energy of the filtered sound is further processed by an EDC and serves as the features for a digital classifier. Experimental results show that the VAD system achieves an averaged 90%/86% speech/non-speech hit rates at 10dB SNR for all tested chips based on a universal classifier trained with data from a single chip, demonstrating the robustness of the proposed acoustic feature extractor against process variations.使用、處理聲學信號的物聯網 (IoT)傳感器及設備現在正占有極大的市場份額。機器學習 (machine learning)技術的進步進一步發展、擴大了物聯網硬件的應用場景和市場。而特徵提取 (feature extraction)的質量很大程度上影響了學習效果的好壞。因爲有限的電源容量，特徵提取的效率極大地影響了許多物聯網設備的工作時限和使用壽命。對一些由薄膜電池或能量收集器等小型電源供電的設備，維持數年的不間斷工作需要硬件功耗被限制在微瓦(μW)内。所以，設計滿足功耗限制的高效率特徵提取電路是十分有必要的。在傳統方法中，聲音信號先被模擬前端直接轉换成數字形式，再由數字信號處理模塊完成特徵提取的任務。此時，數模轉换器 (ADC)、數字信號處理模塊和傳輸模塊需要工作在原始信號的奈奎斯特頻率 (Nyquist-rate)下。因爲許多聲學特徵的帶寬遠小於原始信號的帶寬，同時模擬信號處理電路在較低的信噪比 (SNR)要求下擁有比數字電路更高的效率，先使用模擬電路提取某些特徵再進行數模轉换將會大大降低電路對功耗及帶寬的要求。而特徵數字轉换器 (feature-to-digital converter)就是提取模擬特徵后將其轉换成數字信號的電路。模擬信號處理電路很容易受到工藝偏差 (process variation)的影響，雖然可以通過基於單個芯片的校正或者機器學習改善，但這類方法往往需要很高的成本。所以，本研究的目的是設計受工藝偏差影響較小的超低功耗特徵數字轉换器。本論文將介紹兩個低功耗特徵數字轉换器電路。第一個特徵數字轉换器可以將信號的包絡 (envelope)轉换成數字形式。在許多使用機器學習的人體生物學應用裏，信號的包絡是一個重要的特徵。而該包絡提取轉换電路可以被用來監控人體的呼吸，同時僅僅只有9.6納瓦 (nW)的功耗。該設計使用了基於開關電容 (switched-capacitor)的電路來達到高效和準確的包絡提取。此外，通過一個全新的交替工作 (ping-pong operation)方案，包絡提取電路和數字轉换電路可以直接被整合成一體，從而省去了常見的信號緩衝器 (buffer)且大幅降低了功耗。在測試中，使用該電路提取的人肺音的包絡特徵，可以成功測量使用者的呼吸速率。本文提出的第二個設計是功耗270納瓦的應用於人聲活動檢測 (voice activity detection)的聲音特徵提取器。該設計使用一組無源開關電容帶通濾波器來提取有噪聲干擾下人聲的頻譜特徵。帶通濾波後信號的能量，以包絡的形式由前面提出的包絡數字轉换器轉换爲數字信號，再由數字分類器進行分類。在10dB的信噪比下，該人聲活動檢測系統達到了90%和86%的人聲/非人聲檢測命中率。同時，該結果是在無單獨後期校正的條件下由單一的分類器在多個芯片樣品上取得的。這體現出了該設計對工藝偏差的抵抗力。SHI, Erjia.Ph.D. Chinese University of Hong Kong 2021.Includes bibliographical references (leaves )Abstracts in English and Chinese.Title from PDF title page (viewed on ...