Electronic Thesis and Dissertation Repository

Degree

Master of Engineering Science

Program

Electrical and Computer Engineering

Supervisor

Xianbin Wang

2nd Supervisor

Abdallah Shami

Joint Supervisor

Abstract

With the combination of wireless communications and embedded system, lots of progress has been made in the area of wireless sensor networks (WSNs). The networks have already been widely deployed, due to their self-organization capacity and low-cost advantage. However, there are still some technical challenges needed to be addressed. In the thesis, three algorithms are proposed in improving network energy efficiency, detecting data fault and reducing data redundancy.

The basic principle behind the proposed algorithms is correlation in the data collected by WSNs. The first sensor scheduling algorithm is based on the spatial correlation between neighbor sensor readings. Given the spatial correlation, sensor nodes are clustered into groups. At each time instance, only one node within each group works as group representative, namely, sensing and transmitting sensor data. Sensor nodes take turns to be group representative. Therefore, the energy consumed by other sensor nodes within the same group can be saved.

Due to the continuous nature of the data to be collected, temporal and spatial correlation of sensor data has been exploited to detect the faulty data. By exploitation of temporal correlation, the normal range of upcoming sensor data can be predicted by the historical observations. Based on spatial correlation, weighted neighbor voting can be used to diagnose whether the value of sensor data is reliable. The status of the sensor data, normal or faulty, is decided by the combination of these two proposed detection procedures.

Similar to the sensor scheduling algorithm, the recursive principal component analysis (RPCA) based algorithm has been studied to detect faulty data and aggregate redundant data by exploitation of spatial correlation as well. The R-PCA model is used to process the sensor data, with the help of squared prediction error (SPE) score and cumulative percentage formula. When SPE score of a collected datum is distinctly larger than that of normal data, faults can be detected. The data dimension is reduced according to the calculation result of cumulative percentage formula. All the algorithms are simulated in OPNET or MATLAB based on practical and synthetic datasets. Performances of the proposed algorithms are evaluated in each chapter.


Share

COinS