Pang Jingyue, Liu Datong, Peng Yu, Peng Xiyuan
School of Electrical Engineering and Automation, Harbin Institute of Technology, Harbin 150080, China.
Sensors (Basel). 2018 Mar 24;18(4):967. doi: 10.3390/s18040967.
Effective anomaly detection of sensing data is essential for identifying potential system failures. Because they require no prior knowledge or accumulated labels, and provide uncertainty presentation, the probability prediction methods (e.g., Gaussian process regression (GPR) and relevance vector machine (RVM)) are especially adaptable to perform anomaly detection for sensing series. Generally, one key parameter of prediction models is coverage probability (CP), which controls the judging threshold of the testing sample and is generally set to a default value (e.g., 90% or 95%). There are few criteria to determine the optimal CP for anomaly detection. Therefore, this paper designs a graphic indicator of the receiver operating characteristic curve of prediction interval (ROC-PI) based on the definition of the ROC curve which can depict the trade-off between the PI width and PI coverage probability across a series of cut-off points. Furthermore, the Youden index is modified to assess the performance of different CPs, by the minimization of which the optimal CP is derived by the simulated annealing (SA) algorithm. Experiments conducted on two simulation datasets demonstrate the validity of the proposed method. Especially, an actual case study on sensing series from an on-orbit satellite illustrates its significant performance in practical application.
对传感数据进行有效的异常检测对于识别潜在的系统故障至关重要。由于概率预测方法（例如高斯过程回归（GPR）和相关向量机（RVM））不需要先验知识或累积标签，并且能够提供不确定性表示，因此特别适用于对传感序列进行异常检测。通常，预测模型的一个关键参数是覆盖概率（CP），它控制测试样本的判断阈值，通常设置为默认值（例如90%或95%）。确定用于异常检测的最佳CP的标准很少。因此，本文基于ROC曲线的定义设计了一种预测区间的接收器操作特性曲线的图形指标（ROC-PI），它可以描述在一系列截止点上预测区间宽度和预测区间覆盖概率之间的权衡。此外，对尤登指数进行了修改以评估不同CP的性能，通过最小化该指数，利用模拟退火（SA）算法得出最佳CP。在两个模拟数据集上进行的实验证明了所提方法的有效性。特别是，对在轨卫星传感序列的实际案例研究说明了其在实际应用中的显著性能。
