Abstract:

[Objective] Alarms are important in industrial safety management. The technique to capture the correlation information of alarm variables, especially from historical alarm data, is very beneficial for risk prediction and prevention. But the task is very difficult because some alarms are independent while others are coupled mutually. [Methods] In this paper, a general weight-based multi-state sequential algorithm for correlation analysis is applied for alarm data mining to improve the validity and accuracy of alarm clustering when combined with the traditional agglomerative hierarchical clustering algorithm. To measure the directional relations between alarm variables, this paper proposes a vector correlation concept and use conditional probability modeling to make the alarm correlation among different tags comparable. [Results] The proposed data mining algorithm is shown to be able to find out the vector correlation of alarm variables effectively and correctly when applied in the analysis of power plant alarm data. [Conclusions] Some limitations of the traditional research on alarm variables correlation analysis are solved by the proposed method, which works well on all sequential, non-sequential, regular and irregular alarms. Furthermore, a two-dimensional matrix is used to visualize the vector correlation of alarm variables intuitively and visually.

Key words: alarm, clustering, similarity algorithm, vector correlation