Abstract
As sensors are adopted in almost every field of life, the Internet of Things (IoT) is triggering a massive influx of data. This large amount of data is of little value until it is processed intelligently to extract high-level knowledge which can be used to make decisions. The process of knowledge extraction from data streams is complex predominantly due to heterogeneous data sources, unreliable networks and real-time processing requirements. Different recent studies have showed that solutions based on complex event processing (CEP) have the potential to extract high-level knowledge from these data streams. However, the use of CEP for IoT applications is still in early phase and faces many challenges. First, CEP applications are intended to provide reactive solutions by correlating data streams using predefined rules as the events happen. As the notion of many IoT applications is changing from reactive to proactive where complex events can be predicted before they actually happen, solutions based on CEP required an extension to address this issue. To this end, this work proposes a proactive method based on CEP and machine learning (ML) where historical data is exploited using ML part and combined with real-time flow of CEP to provide the basis for predictive event processing. Second, systems based on CEP deploy static rules and there is no means to update the rules according to the current context automatically. In order to address this issue, this thesis proposed a novel method based on ML to find CEP rules automatically and update them according to the current context. Third, in state-of-the-art CEP systems, events are correlated using absolute rules where a complex event detected is either true or false. Given the sporadic nature of IoT, missing and uncertain data is a common phenomenon and CEP systems of today are unable to take this inherent uncertainty of real-world events into account while taking decisions. This thesis addressed this issue by proposing a probabilistic event processing approach by extending state-of-the-art CEP by combining it with Bayesian networks (BNs). Finally, the size and complexity of the IoT data presents a generic challenge which is being addressed throughout in this thesis. The above mentioned contributions were evaluated using real-world data collected from heterogeneous sources to prove the accuracy and reliability of the proposed methods. The feasibility and applicability of the proposed solutions were demonstrated by implementing it for real-world applications. The work presented in this thesis significantly improves state-of-the-art methods and provides a fundamental building block towards extracting knowledge from raw IoT data streams.