Resource management in sensing services with audio applications

Resource management;Sensing services;Internet of Things;Audio classification;Guided-processing;Feature-sharing

Le, Long Nguyen Thang

关键词: Resource management;    Sensing services;    Internet of Things;    Audio classification;    Guided-processing;    Feature-sharing;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/97414/LE-DISSERTATION-2017.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Middleware abstractions, or services, that can bridge the gap between the increasingly pervasive sensors and the sophisticated inference applications exist, but they lack the necessary resource-awareness to support high data-rate sensing modalities such as audio/video. This work therefore investigates the resource management problem in sensing services, with application in audio sensing. First, a modular, data-centric architecture is proposed as the framework within which optimal resource management is studied. Next, the guided-processing principle is proposed to achieve optimized trade-off between resource (energy) and (inference) performance.On cascade-based systems, empirical results show that the proposed approach significantly improves the detection performance (up to 1.7x and 4x reduction in false-alarm and miss rate, respectively) for the same energy consumption, when compared to the duty-cycling approach. Furthermore, the guided-processing approach is also generalizable to graph-based systems. Resource-efficiency in the multiple-application setting is achieved through the feature-sharing principle. Once applied, the method results in a system that can achieve 9x resource saving and 1.43x improvement in detection performance in an example application.Based on the encouraging results above, a prototype audio sensing service is built for demonstration. An interference-robust audio classification technique with limited training data would prove valuable within the service, so a novel algorithm with the desired properties is proposed. The technique combines AI-gram time-frequency representation and multidimensional dynamic time warping, and it outperforms the state-of-the-art using the prominent-region-based approach across a wide range of (synthetic, both stationary and transient) interference types and signal-to-interference ratios, and also on field recordings (with areas under the receiver operating characteristic and precision-recall curves being 91% and 87%, respectively).