Behavioral pattern discovery from sensor data stream

The advances in technologies and microelectronic devices have led to the development of sensor nodes which are able to sense, process and transmit data, and co-ordinate among themselves to form wireless sensor networks (WSNs). Sensors embedded in objects can sense and interact with the environment around us to achieve ubiquitous computing. The deployment of large-scale WSNs is increasing day-by-day in many application domains such as habitat monitoring, object tracking, environment monitoring, military, disaster management, as well as smart environments. These applications generate huge volume of dynamic, geographically distributed and heterogeneous data in the form of streams. Such data are useless unless mined accurately and efficiently for useful knowledge. Data mining techniques play an important role to efficiently extract and analyze usable information from the raw data to facilitate automated or human induced decision. Data stream from WSN can be mined to extract knowledge in real time about the sensed environment (e.g., mining certain behaviors) and the network itself (e.g., predicting faulty nodes), but all these present new challenges for data mining techniques since traditional data mining techniques are not directly usable because of resource constraints in WSN and limitation of current techniques. Given the fact that WSNs will be the integral parts of the future Internet of Things (IoT) environment, there is a pressing need to develop efficient techniques to mine knowledge from the sensor data patterns in order to make prompt intelligent decisions in an integrated service architecture. The proposed thesis focuses on the design of a framework for extracting behavioral patterns that capture the temporal relations among sensor nodes in a WSN in an efficient way. Three types of behavioral patterns are introduced: Associated sensor patterns (ASPs), Share-Frequent Sensor Patterns (SFSPs) and Regularly Frequent Sensor Patterns (RFSPs). The formal definition of the required knowledge, the data structure to store the data, and the mining techniques that relate to the extraction of the behavioral patterns are defined. ASPs capture association-like-co-occurrences as well as temporal correlations which are linked with such co-occurrences from sensor data. To capture such patterns from static sensor dataset as well as sensor data stream, two compact tree structures called ASP-tree and sliding window ASP-tree (SWASP-tree) have been proposed which use pattern growth-based mining technique to mine patterns with only one scan over dataset. Both trees maintain `build once and mine many' property which make them highly suitable for interactive mining. SFSP discover share relations among the sensors considering the number of triggers rather than binary pattern frequency and builds a highly compact tree structure called share-frequent sensor pattern tree (ShrFSP-tree). ShrFSP-tree has been enhanced into parallel ShrFSP-tree (PShrFSP-tree) for faster mining in parallel and distributed environment and demonstrates better time and memory efficiency than the existing most efficient algorithms. RFSPs discover the shape of occurrence behaviors among sensors using a tree structure called RFSP-tree. To make RFSP more efficient and integrate in the Hadoop platform, a MapReduce-based framework is proposed that mines RFSPs among a set of dedicated nodes in parallel and provides better service availability and timely response of mining results. Knowledge extracted from these patterns will be useful to improve the performance of WSN and the quality of services (QoS) it provides in a number of ways, e.g., real-time knowledge extraction about the physical environment, identification of chain of correlated events in industries as well as resource management and operation in WSN