Design of an ultra-low-power always-on analog front-end for audio events detection in forest

The constant advances in integrated artificial intelligence (AI), low power electronics and increased network agility enable the deployment of connected smart sensor nodes exchanging data over the internet, also known as the Internet of Things (IoT). This extensive deployment raises ethical and environmental concerns. Sensor materials are rare and sometimes difficult to harvest. Besides, the e-waste generated by the replacement of each sensor is harmful to the environment. Each sensor needs to be autonomous, non-invasive and exhibit a long lifespan to prevent such issues. If design efforts are made to meet these constraints, the IoT could be profitable in ecosystem monitoring. Among these ecosystems, forests provide essential services and represent resources to protect in the context of climate change. This thesis is part of a larger project aimed at developing an automated acoustic sensor node monitoring several parameters of forest conditions. Because relevant sounds can happen at any time, an always-listening sensor is required. Nevertheless, keeping the entire sensor always awake is not a possibility due to the restricted power budget. For that reason, the purpose of this master thesis is to achieve the design of an ultra-low-power (ULP) analog-front-end (AFE), enabling the second stage of the sensing system when any relevant sound is detected. First, a low-power MEMS microphone converts the acoustic wave into an electrical signal. It features -37 dBV sensitivity and a bandwidth spanning from 90 Hz to 16 kHz. Second, from the microphone specifications, constraints are set for the low-noise amplifier (LNA), which consumes 530 nW and provides a constant 30 dB gain across the microphone bandwidth. Afterward, the energy of the amplified signal is extracted by a 420 nW half-wave rectifier followed by a passive low-pass filter. Finally, the second stage of the sensing system is enabled whether the comparator at the end of the chain evaluates the signal energy as sufficient with respect to its comparison voltage. This last circuit consumes 66 nW and presents 4 mV hysteresis to prevent its binary output voltage from instability. The comparator is also provided with adjustable comparison voltages to avoid always toggle when the background noise increases. In this way, the AFE can adapt to distinct background noise levels. The aforementioned analog circuits use a single supply voltage of 600 mV. All their transistors operate in weak inversion for low power considerations. The always-listening AFE (or always-on chain) senses sounds as low as 37 dBSPL (minimum level of detection). It can enable the rest of the sensor for sound levels greater than 57 dBSPL (minimum toggle level) and adjust this level depending on the background noise. The total power consumption of the proposed design is 1.02 μW, which is 437 times less than the previous AFE design proposed in the project on acoustic monitoring in forests. As a result, the always-on chain aligns with state-of-the-art voice activity detection (VAD) systems for the considered bandwidth. The main limitations of the proposed design are associated with the variation of performance with PVT corners and statistical variability. This will be thoroughly discussed in this work. The design presented in this thesis exhibits promising results in the acoustic monitoring of forests thanks to its low power consumption and high bandwidth. If subsequent research follows the footsteps of this thesis, forests could be effectively monitored by ULP sensors in the near future, hence protecting their resources and limiting the impact of their degradation on the environment.