Context Prediction based Collaborative IoT Architecture for Smart Environment

Abstract

The advancement in sensing and communication technologies keeps adding new dimensions to the world of information and communication technologies (ICTs) i.e. connectivity from anyplace, anytime, to anything. This hyper-connectivity leads to a new technological paradigm called Internet of Things (IoT) which is drastically transforming our commercial, social, and personal sphere. The lack of standard procedure for communication among the constituent heterogeneous parts(devices) of IoT post a number of research challenges like connectivity, data processing, privacy preservation etc. Many research studies have been dedicated to address the communication architecture and data processing problem for these things. These studies resulted in various models which enable connecting things, processing the collected data and proactively responding to the situation. Some of these models are application-specific while others introduce a new terminology i.e. everything as a service (XaaS) into existing Service Oriented Architecture (SOA). However, a purely SOA based service oriented system may be wasting resources on processing unnecessary data. Moreover, the concerns over privacy and data protection in a SOA based systems are widespread, particularly as sensors and smart tags can track users‟ movements, habits and ongoing preferences. To the best of our knowledge, The state of the art studies are application specific or domain specific. The scope of existing state of the art models is limited to provide solution for a specific application area. To extend the IoT functionality to an interoperable domain problem, a cross-domain architecture for IoT that will bridges gap between the service oriented paradigm and standard IoT concepts, is presented in this dissertation. The gap has been bridged by dividing the data processing between the low power & low processing capability devices and the cloud based servers. Performance of proposed architecture is evaluated by integrating applications from three different domains. First application is a sophisticated collaborative context prediction approach based on inductive learning. Experiments are conducted to demonstrate the effectiveness of proposed techniques on real world data sets and virtual data sets. The results show that proposed approach can be applied effectively to diverse application areas. Second application is forecasting electricity consumption in multi-family residential buildings. Current state of the art approaches for energy management in residential buildings use traditional forecasting techniques based on statistical analysis and machine learning approaches applied on monthly utility meters data, thus restricting hourly forecast. The smart metering infrastructure enables collection of the electricity consumption data at a fine temporal level. Therefore, we applied well recognized machine learning techniques i.e. neural network, regression tree and genetic programming on smart meters data to forecast electricity consumption in residential buildings. The third application area is multimedia content recommendation in smart home environment. The home network based model is proposed for this application which proactively responds to the user behavior in a smart home. A state of the art learning approach i.e. deep learning is used to enhance the performance of content based collaborative recommendation system. For this purpose, a hierarchical Bayesian model based deep learning is used, which couples a Bayesian formulation of the stacked denoising autoencoder and probabilistic matrix factorization. Experiments on real-world datasets shows that this model can improve the performance of the content recommendation system.