Abnormal data detection for multivariate alarm systems based on correlation directions

- A method is proposed to detect abnormal data segments for multivariate time series.
- Correlation directions are exploited as the features to detect abnormal conditions.
- Key turning points are determined by a piecewise linear representation of segments.
- Spearman's rank correlation coefficients are calculated for correlation directions.
- Numerical and industrial examples are provided as illustrations.

This paper proposes a method to detect abnormal data segments from historical multivariate time series, which are common prerequisites for rationalization of industrial alarm systems. Correlation directions among process variables are taken as the features to detect abnormal conditions. To find time instants of changing correlation directions, key turning points (KTPs) are determined by a piecewise linear representation of multivariate time series. Correlation directions in each data segment between adjacent KTPs are calculated from Spearman's rank correlation coefficients and associated hypothesis tests. Data segments are classified into normal or abnormal ones by comparing the calculated correlation directions with their counterparts in normal conditions obtained from process knowledge. Numerical and industrial examples are provided to illustrate the proposed method.