Study On Outlier Detection Techniques Based On Multiple Attributes In Wireless Sensor Networks

Wireless sensor network (WSN) are applied in various fields such as environment monitoring and traffic controlling. As one of the most important applications, outlier detection becomes more and more popular. However, because of the traits such as low accuracy, limited hardware resources, fragile anti-disturbance, sensor nodes would produce errors as a result of environment noise, hardware disturbance, environment changes, lost of energy and hardware failure. Data with errors will have a disgusting influence on accuracy of the detection results, which waste lots of human and material resources. So an in-network detection technique for outlier detection with high accuracy is very desirable.By analyzing attribute similarity of sensor data, the concept of correlating attributes is established. Based on this, multidimensional data space is designed. The true outlier is determined through comparison of the similarity of different data. The attributes variations are proposed to measure the variations when some correlating attributes change. Outlier detecting process contains the time correlation detection and space correlation detection. For time detection process, some historical data are saved in every node, the integrated similarity is calculated between current data and historical data by weighted method, and then the average similarity is obtained. If the average similarity is less than the threshold, the current data is non-normal data. The node which sensed the data transmits outlier message to neighbor. For space detection process, the neighbors calculated the data similarity with the temporality node. If the similarity is more than the threshold, the current data is considered outlier; the node should transmit the data to the base and send ACK to the temporality node. If the data is noise, the neighbors are used instead of current data. The data window moves forward when the current data is logged. Based on historical data, the user can mine the correlating attributes information and the rules of outlier in sink, then send the results to the node. So the detection may be more exact and consume less energy.Finally, experiments are carried out to prove the proposed techniques' feasibility, validity, and real-time nature. Extensive experiments show that the proposed MATSDT can detect the outliers, clean the noisy, and reduce the uploading number of data.