Exploiting the wisdom of crowds in wireless networks : sensing, labelling, and data aggregation

Abstract

Driven by the unprecedented growth of Internet of Things (IoT) and advanced 5G communication techniques, billions of wireless IoT devices are deployed to collect data for a wide range of applications, such as public-safety surveillance, traffic monitoring, and artificial intelligence (AI). The traditional wireless sensor networks have difficulty in supporting next-generation sensing tasks requiring ubiquitous coverage and a massive number of sensors. Leveraging the sensors on available billions of mobile devices as well as their users’ data and intelligence presents a novel solution called mobile crowd sensing (MCS), which promises large coverage, high scalability, low cost of deployment, and easy maintenance. Achieving these promised gains requires the effective designs of data exploitation process including data sensing, aggregation, and analytics. This dissertation contributes to the emerging area by investigating the incentive mechanism for crowd data sensing, Over-the-Air Computation (AirComp) for data aggregation, and crowd labelling for supporting the data analytics.

First, to incentivize human participation in MCS and prolong battery lives of sensing devices, this dissertation presents a novel design of wirelessly powered MCS system by applying wireless power transfer (WPT) as the incentive mechanism. The power allocation, sensing data size, sensor transmission time, and compression ratio are jointly optimized to maximize data utility and minimize energy consumption. A crowd-sensing priority function of the above variables is derived to be the metric for resource allocation over devices based on their users’ biding profiles. By analysis, it is found that when the sensing-data size is sufficiently large, a high compression ratio is preferred for energy saving.

Next, to accelerate the wireless data aggregation (WDA) over the dense IoT devices, AirComp has been proposed to merge computing and communication by exploiting analog-wave addition in the air. In this dissertation, we advocate the integration of AirComp and WPT to not only power the sensing devices for AirComp, but also reduce the computation error by jointly optimizing the wireless power control, energy and data aggregation beamforming. The results indicate that the optimal energy beams should point to the dominant Eigen-directions of the WPT channels, and the optimal power allocation attempts to equalize the multiple cascaded WPT-AirComp channels to reduce the AirComp error.

Last, we investigate the application of MCS on a typical data analytics scenario, i.e., AI model training via supervised learning, which entails labelling of the enormous data prior to the training process. To tackle this challenge, we explore a new perspective of wireless crowd labelling (WCL) that can multicast raw data to imperfect mobile annotators for repetition labelling. To maximize the labelling throughput, defined as the number of labelled objects under a constraint on labelling accuracy, this dissertation presents a framework of joint optimization of the encoding rate, annotator clustering, and sub-channel allocation. The scheme of sequential clustering of annotators arranged in the order of decreasing signal-to-noise ratios is proved to be optimal. The result substantially simplifies the optimization problem and reduces the optimal policy computation to a tree search.