21.6 A 12nW always-on acoustic sensing and object recognition microsystem using frequency-domain feature extraction and SVM classification

Abstract

IoT devices are becoming increasingly intelligent and context-aware. Sound is an attractive sensory modality because it is information-rich but not as computationally demanding as alternatives such as vision. New applications of ultra-low power (ULP), ‘always-on’ intelligent acoustic sensing includes agricultural monitoring to detect pests or precipitation, infrastructure health tracking to recognize acoustic symptoms, and security/safety monitoring to identify dangerous conditions. A major impediment for the adoption of always-on, context-aware sensing is power consumption, particularly for ultra-small IoT devices requiring long-term operation without battery replacement. To sustain operation with a 1mm2 solar cell in ambient light (100lux) or achieve a lifetime of 10 years using a button cell battery (2mAh), <20nW power consumption must be achieved, which is more than 2 orders of magnitude lower than current state-of-the-art acoustic sensing systems [1,2]. More broadly a previous ULP signal acquisition IC [3] consumes just 3nW while 64nW ECG monitoring system [4] includes back-end classification, however there are no sub-20nW complete sensing systems with both analog frontend and digital backend.