Abstract: The historical normal and abnormal data sets of process variables are premises of assessing and optimizing alarm performance and designing dynamic alarm trip-points of industrial alarm system. This paper proposes an improved abnormal data detection method, which is based on the correlation coefficients between process variables. The main idea is to divide multivariate time series of process variables on the basis of correlation coefficient values and suitable length of data segments, obtain the mutual variation directions of process variables through Spearman rank correlation coefficient and corresponding hypothesis test, and detect abnormal data segments that have inconsistent variation directions with prior knowledge for normal conditions. Simulation examples and industrial case are provided to validate this method.

Key words:  Industrial alarm system, variation direction, correlation, multivariate time series

摘要： 工业报警系统普遍存在报警过多的问题，为了获取过程变量的正常与异常历史数据段，从而设计工业报警系统的动态报警阈值，进行性能评估和优化，提出了一种基于过程变量之间的相关系数进行异常数据段检测的改进方法。该检测方法将相关系数和数据段长度作为分段依据，对过程变量的多元时间序列进行分段，采用Spearman秩相关系数获取变量之间的相关关系，从而检测出不符合正常趋势的异常数据段。仿真算例和工业案例表明该方法可真实地反映过程变量之间的相关关系，并准确检测出所存在的异常数据段。

关键词: 工业报警系统, 异常数据段检测, 相关系数, 多元时间序列