Multiple-Output-Gaussian-Process Regression-Based Anomaly Detection for Multivariate Monitoring Series
Jingyue Pang, Datong Liu, Peng Yu, Xiyuan Peng · 2018
Compared with the anomaly detection for univariate series, it is more challenging to detect the anomalies within the multi-sensors timely and effectively, especially in the aerospace area where the multi-sensors have the large-scale and complex relation. As the prediction models have strong explanatory and online applicability, they have been widely applied for detecting anomalies within multiple monitoring series. Specially, Gaussian process regression (GPR), a probability prognostics model, which can provide uncertainty presentation is the focus of this work. However, the anomaly detection for multi-sensors is generally realized by modelling some independent GPRs, and the data points beyond the prediction interval are labeled anomalies. This detection framework neglects the relationship between each output, and it will cause the decrease of the detection accuracy. Therefore, given the strong relation of multiple monitoring series, this paper proposes a multiple-output-gaussian-process regression (MOGP) -based method for realizing the anomaly detection of multisensors, where SPE statistics is introduced to compute the anomaly score. Some experiments on real monitoring series validate its superiority and convince its strong applicability in the actual system.