Data-driven Techniques on Alarm System Analysis and Improvement

Abstract

To meet the demands of safety, quality and efficiency, process monitoring is of great importance. However, a serious problem exists in the industry: too many alarms are raised for operators to handle. Consequently, techniques need to be developed in order to reduce nuisance and false alarms to an acceptable level. Motivated by this, my thesis focuses on alarm systems improvement, specifically the development of data-driven techniques for the analysis and design of alarm systems. Developed methods are based on either process data or alarm data, the two types of data mainly used in alarm systems. Three problems are considered. First, a univariate alarm signal filtering technique is discussed. The design of an optimal alarm filter for the best alarm accuracy, namely, minimizing a weighted sum of false and missed alarm rates (probabilities), is presented. Moreover, a sufficient condition for moving average filters being optimal linear alarm filters is also provided. Second, alarm flood pattern analysis based on multivariate alarm data is addressed. A modified Smith-Waterman algorithm considering time stamp information is proposed for alarm flood pattern matching. Third, the application of a new multivariate statistical analysis technique, the principal component pursuit (PCP) method, to process monitoring is thoroughly discussed. An optimal scaling method is proposed as the preprocessing step. A coordinate descent algorithm is provided to search for the optimal scaling vector, whose global convergence is proved. After multivariate process modeling, a PCP-based fault detection and diagnosis approach is introduced. An industrial project on a major extraction process in Alberta to improve its alarm system is described. Based on the historical data, a predicted reduction on the alarm rate by applying a variety of alarm system rationalization techniques is estimated. A significant improvement on the alarm system is expected.