Inferring Missing Sensor Values and Recommending Sensor Deployment Locations Based on Urban Big Data: A Case Study on Air Quality

• Main Authors: Hsun-Ping Hsieh, 解巽評
• Other Authors: Shou-De Lin
• Format: Others
• Language: en_US
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/qw94dv

博士 === 國立臺灣大學 === 資訊網路與多媒體研究所 === 103 === With the gradual maturity of the persuasive techniques for networking, sensors, and Internet, the paradigm of information and communication technology has shifted from fundamental facilities and theories to real-world ubiquitous applications, such as urban computing, sensor networks, and smart cities. Along with the power of big data techniques, such advances also bring new research opportunities and practical problems, especially for Internet of Things (IoT). This thesis aims to answer two questions. First, how to infer real-time sensor value of any arbitrary location given the environmental data and the sensor data collected from monitoring stations that are extremely sparse in urban areas. Second, if the government agency needs to deploy a number of new sensors to improve the quality of sensor value inference, how to automatically learn and determine the best locations to fulfill such purpose? These two problems are very essential for urban computing and big sensor data analysis because sensors will be mounted in urban areas in the future era of Internet of Things, and are considerably challenging since for most of the locations (more than 99%) we do not have any sensor data to train a model from. In this thesis, by considering air quality indices as sensor values and treating air-quality monitoring stations as sensors, we aim to solve the two research problems in an effective, efficient, and scalable manner. We develop a general-purpose semi-supervised inference model, which is capable of not only intelligently learning to correlation between air quality values and heterogeneous spatial and temporal features of city dynamics, including meteorology, human mobility, structure of road networks, and point of interests (POIs), but also accurately predicting the air quality values of arbitrary locations without monitoring stations placed ever. In addition, to facilitate the cost-effective deployment of new sensors, we devise an entropy-minimization model to efficiently recommend the geographical locations such that new sensors established there can lead to the maximum improvement of air quality inference accuracy. We evaluate the proposed models using a huge-scale air quality data in Beijing city. Experimental results exhibit a set of clear advantages over a series of state-of-the-art and commonly used methods for both tasks, and suggest the models have superior potential for applications for current big sensor data analysis and the upcoming era of Internet of Things.