基于空-时近邻与似然比检验的传感器网络异常点检测
Outlier detection based on spatio-temporal nearest neighbors and a likelihood ratio test for sensor networks

针对传感器网络中由于传感器故障造成的异常点检测问题，该文提出一种基于传感器与其空-时近邻点在测量数据之间的差异，采用似然比检验来判断传感器是否故障的异常点检测方法。在空间维，该方法基于最大后验概率选取待检测传感器当前时刻的空间近邻点；在时间维，该方法选取待检测传感器在之前若干个时刻的测量值作为其时间近邻点。然后根据待检传感器与其空-时近邻点测量数据之间的差异对其异常程度进行量化，并采用似然比检验判断待检测传感器是否故障。结果表明：该方法与已有的异常点检测方法相比，在相同的虚警率下取得了更高的检测率。例如在虚警率为10%时，该方法将检测率提升了10%~30%。
Abstract：A spatio-temporal nearest neighbors and likelihood ratio test method was developed to detect outliers caused by sensor failures in sensor networks. In the space dimension, a sensor's spatially nearest neighbors were selected using a maximum posterior probability criterion while in the time dimension, the temporal nearest neighbors were previous observations from the same sensor. Each sensor's reading was evaluated based on differences between its earlier measurements and those of its neighbors with a sensor failure model and likelihood ratio test used to detect whether the sensor had failed. Tests show that this approach gives a higher detection rate for the same false alarm rate than existing outlier detection approaches. For example, for a 10% false alarm rate, the detection rate was increased by 10%-30%.