Master thesis defense Martin Braquet Supervisors: D. Bol and R. Sadre Master in Electromechanical Engineering 25 June 2020
Internet of Things (IoT): billions of connected smart sensors → Currently not environmentally sustainable • Beforehand: pressure on critical • Afterwards: ecotoxicity of e-waste elements (exported, incinerated, landfills) 25/06/2020 Martin Braquet - Master thesis defense 2
Ecosystem destruction: climate chang e → ec osystem monitoring • Focus on monitoring in forest • water and soil conservation [1] • genetic resources for plants and animals • source of wood supply • benefits on human physical and mental health [2] Current approach: manual and sporadic sampling requiring human presence → evolution characterization limited by low observation frequency 25/06/2020 Martin Braquet - Master thesis defense 3
Autonomous and efficient audio smart sensor for bird monitoring Use case • Energy harvesting from environment • Environmentally-friendly and non-toxic components • Audio signal processing for bird classification • LPWAN communication • Data transmission • Reconfiguation for firmware updates 25/06/2020 Martin Braquet - Master thesis defense 4
• Challenges and requirements • Design • Energy storage • Sensing • Power management • Solar cells and supercapacitor • Validation • Model view • Power budget and MPPT • Inference for bird monitoring • Algorithm • Live demo • Conclusion and outlook 25/06/2020 Martin Braquet - Master thesis defense 5
Supercapacitors : good trade-off between High service life • capacitors: high power density → fast (dis)charge • rechargeable batteries: high energy density → reduced volume But high leakage current 25/06/2020 Martin Braquet - Master thesis defense 6
Critical and toxic elements overused in electronics: lithium, cobalt, gold, silver, ... Li-ion batteries Supercapacitors Electrodes lithium, nickel, cobalt, graphite porous carbon (graphite) Electrolyte lithium salts liquid salts (lithium) → Selection for this work: supercapacitors (far less toxic and resourse-intensive) 25/06/2020 Martin Braquet - Master thesis defense 7
Physical stimulus: sound wave (in dB or dBSPL) • Transduced with a microphone • 50m detection • Bird frequency range (1kHz - 8kHz) Current Voltage Sound pressure 25/06/2020 Martin Braquet - Master thesis defense 8
IoT applications: condenser microphones (MEMS or electret ) • Figures of merit Power consumption Self-noise Very low noise (with same power consumption) 25/06/2020 Martin Braquet - Master thesis defense 9
Analog front-end Sound • Transimpedance amplifier Microphone Waves (AC) Biasing (DC) • Input/output matching Output voltage range (0 - VCC) Input pressure range (16 – 61 dBSPL) Op amp biasing @ VCC/2 25/06/2020 Martin Braquet - Master thesis defense 10
HF pole: 20kHz LF pole: 20Hz Analog front-end • Transimpedance amplifier • Input/output matching • Pole placement LF pole: 5Hz 25/06/2020 Martin Braquet - Master thesis defense 11
Analog front-end • Noise • Thermal noise (resistors) • Op amp input-refered current noise • Op amp input-refered voltage noise • Microphone self-noise 25/06/2020 Martin Braquet - Master thesis defense 12
Analog front-end • Noise • Thermal noise (resistors) Minimum input pressure • Op amp input current noise • Op amp input voltage noise • Microphone self-noise Selected op amp Pareto front → trade-off noise/power consumption for op amp selection 25/06/2020 Martin Braquet - Master thesis defense 13
Power management unit from e-peas • Energy • Harvesting : solar cells (with MPPT) 2.5V supply voltage • Storage : supercapacitor • Low-dropout (LDO) regulation • with low quiescent current 25/06/2020 Martin Braquet - Master thesis defense 14
Current budget (Power budget 𝑄 𝑐𝑏𝑢𝑢 = 𝑊 𝑐𝑏𝑢𝑢 𝐽 𝑢𝑝𝑢 depends on supercap voltage since LDO in PMU) • Sensing • Microphone biasing • Op-amp supply • Power management • Supercap leakage • PMU quiescent current • Data processing (MCU/RF) • Alternance run / sleep modes with limited duty cycle (1/3) 25/06/2020 Martin Braquet - Master thesis defense 15
Maximum Power Point Tracking Solar cells • Voltage adaptation for maximum power harvesting • Maximization of harvested power per unit area IV curve of the KXOB25-14X1F at Standard Condition: 1 sun (= 100 mW/cm2), Air Mass 1.5, 25°C (from datasheet) 25/06/2020 Martin Braquet - Master thesis defense 16
Solar cells • Luminosity profile along the day • → Estimation of harvested power Daily luminosity In a shady place (Louvain-la-Neuve, from March 3 to March 6, 2020) 25/06/2020 Martin Braquet - Master thesis defense 17
Summary Input power (1 cell) Output current MCU in operation only during the day (for power reduction) 25/06/2020 Martin Braquet - Master thesis defense 18
6 solar cells required MCU MCU MCU MCU Sun Sun Sun Sun 25/06/2020 Martin Braquet - Master thesis defense 19
90F/4.2V supercapacitor required • Below the 4.5V PMU limit • Avoids high voltage → keep high LDO efficiency 25/06/2020 Martin Braquet - Master thesis defense 20
• CAD • Real MCU/RF 6 solar cells Analog front-end Power management unit Dimensions: 143 mm × 82 mm × 25 mm Microphone Supercapacitor 25/06/2020 Martin Braquet - Master thesis defense 21
• Power budget Experimental current Theoretical current Supercap leakage 90 µA 500 µA Sensing 904 µA 536 µA Power management unit 0.5 µA 0.6 µA *MCU current consumption (run/sleep Microcontroller 3.88 mA* (measured experimentally) duty cycle of 1/3) for • ADC sampling @ 20kHz • FFT operations (N = 128) Total 4.874 mA 4.916 mA • Bird inference • Maximum power point tracking 25/06/2020 Martin Braquet - Master thesis defense 22
Bird species discrimination inside the microcontroller • Among 4 common species in Europe: pigeon, blackbird, great tit and blue tit • Limitations: memory, speed (≈ power) • Fast Fourier transform (FFT): spectral domain • Use of spectrogram • Machine learning approach (KNN) 25/06/2020 Martin Braquet - Master thesis defense 23
Spectrogram : time-frequency representation of audio signals Trade-off time / frequency resolution • FFT size: 128 • Sampling frequency: 20 kHz Frequency resolution: 156 Hz Post-processed spectrogram On-chip spectrogram (4 great tit songs at regular intervals) (One great tit song) 25/06/2020 Martin Braquet - Master thesis defense 24
Different frequency range for each bird ≈ 2 kHz ≈ 1 kHz ≈ 6 kHz ≈ 4 kHz 25/06/2020 Martin Braquet - Master thesis defense 25
Machine learning approach • Feature extraction: weighted average frequency Mean amplitude @ f[I] • Mean frequency of the whole spectrogram • Feature selection • Only one feature for reduced complexity • Inference • 𝑙 -nearest neighbors (KNN) algorithm ( 𝑙 = 5 ) • Learning phase • 6 audio samples per species 25/06/2020 Martin Braquet - Master thesis defense 26
Machine learning approach On learning • Validation phase samples 94% recovered On new test All recovered samples (but small test set) (3 per species) 25/06/2020 Martin Braquet - Master thesis defense 27
Live demo • Bird classification • Sound generated from laptop speakers • Pigeon • Blackbird • Great tit • Blue tit 25/06/2020 Martin Braquet - Master thesis defense 28
Live demo • Backup video 25/06/2020 Martin Braquet - Master thesis defense 29
Ultra-low-power energy-harvesting audio sensor for ecosystem monitoring • Fully autonomous and sustainable Electret microphone • Context of resource saturation in IoT • Amplification circuit: trade-off noise / power in op-amp Solar cells Important demand for forest monitoring • • MPPT optimization Context of climate change • Daily luminosity estimation Bird classification Supercapacitor • Spectrogram: trade-off time / frequency resolution • • Less toxic and resource-intensive KNN algorithm: simple but fast (low power) • Good trade-off power / energy density 25/06/2020 Martin Braquet - Master thesis defense 30
SWOT analysis Important numbers • Lifetime: > 15 years • Supply voltage: 2.5 V • Average power: 20 mW • Input referred noise: 14.22 dBSPL • Sound pressure range: 16 – 61 dBSPL • Sound frequency range: 20 – 20k Hz • Detection duty cycle: 1/3 (during the day) • Classification: 4 birds ( 94% accuracy) 25/06/2020 Martin Braquet - Master thesis defense 31
Custom integrated circuit for ML inference in hardware Audio smart sensors Badami, 2015 [5] This work Pham, 2014 [3] Zhao, 2012 [4] Power supply Supercap + solar Rechargeable batteries Rechargeable batteries / cells (AA) 6 µW Power 20 mW 330 mW 73 mW consumption / Not autonoumous Manual recharge Never Every night Every week 2 kHz Sampling rate 20 kHz 8 kHz 8 kHz 640 Hz (feature extraction) CPU frequency 32 MHz 47.5 MHz 48 MHz Microphone Electret MEMS Electret Passive Use case Bird monitoring Audio streaming Audio surveillance Voice Activity Detection → Low sampling rate not for HF bird songs → Power consumption only suited for batteries 25/06/2020 Martin Braquet - Master thesis defense 32
Recommend
More recommend
