Anomaly detection based on uncertainty fusion for univariate monitoring series

Detecting the anomalies timely in the condition monitoring data, which are highly relevant to the potential system faults, has become a research focus in many domains. Among the various detection methods available, the prediction-based algorithms are popular without using prior knowledge and expert labels. Additionally, these methods can take the time-ordered specialty into account which is highly significant for time-series-based anomaly detection. However, the detected feedback is binary, especially, due to the influence of inaccurate confidence interval (CI), the false alarm phenomena occur frequently based on predicted models. Thus this paper proposes an Uncertainty Fusion method to realize anomaly detection. Firstly, in order to estimate the data uncertainty, the Gaussian Process Regression (GPR) is applied to perform the prediction with uncertainty presentation. Then, based on the GPR model, the improved k-fold cross-validation is combined to represent the model uncertainty. Moreover, the quantitative error index is designed to provide more detecting information for decision-making. Eventually, the effectiveness of the proposed method are verified by different simulated and open-source data sets, as well as the real application in mobile traffic data detecting. The quantitative results on simulated data experiments show the proposed method can largely eliminate the false alarms without sacrificing much detection rate compared with the basic GPR model. Especially, the experiments on periodic data sets with high Signal Noise Ratio have the better performance. And the mobile traffic data detecting proves the Uncertainty Fusion method can expand the basic GPR model to meet the real industrial requirements.