(587a) Probabilistic Model-Based Method of Improving Alarm-System Performance

Automatic alarm systems are inseparable components of current industrial plants. They are designed such that in case of abnormal events operators are notified and the appropriate precautionary actions are taken in shortest possible time. However, in many cases the number of less informative alarms annunciated is much larger than it could be useful or even be handled. This situation is usually called “alarm flood”. On the other hand, alarm thresholds should not be set such that the alarm system fails to announce the occurrence of an abnormal event in an effective way [1].

To address these issues, the design and implementation of efficient alarm systems have been an active research area. The goal has been to decrease the number of false alarms without increasing the number of missed alarms. To this end, a variety of techniques have been proposed. These techniques include multivariate monitoring [2], model-based monitoring [3, 4], threshold optimization [5], and alarm signal processing [6].

This paper presents a probabilistic model-based method of improving alarm-system performance. The performance is quantified in terms of number of false and missed alarms. This method uses a novel data-driven probabilistic modeling approach to calculate the differences between measurements and model-predictions of process variables. The model is trained properly to cover all possible states of the process using the process historical data. The calculated differences will be used to identify abnormal situations and trigger appropriate alarms. It is shown that the modeling approach can model highly non-linear and non-monotonic behavior observed in data with minimum number of parameters. To make the calculation of the differences computationally tractable, we use an efficient probabilistic inference technique, which in combination with the model presents a unique framework for probabilistic reasoning particularly when there are a large number of variables to deal with.

References

[1] Rothenberg, D. H. Alarm Management for Process Control: A Best-Practice Guide for Design, Implementation, and Use of Industrial Alarm Systems. New York: Momentum Press, 2009.

[2] Chen, T. On reducing false alarms in multivariate statistical process control. Chemical Engineering Research and Design 2010, 88(4), 430-436.

[3] Mohseni Ahooyi, T.; Arbogast, J. E.; Oktem, U.; Seider, W. D.; Soroush, M. Maximum-Likelihood Maximum-Entropy Estimation of Multivariate Probability Density Functions. AIChE J. 2014, 60(3), 1013−1026.

[4] Mohseni Ahooyi, T.; Arbogast, J. E.; Oktem, U.; Seider, W. D.; Soroush, M. Estimation of Complete Discrete Multivariate Probability Distributions from Scarce Data with Application to Risk Assessment and Fault Detection. Industrial & Engineering Chemistry Research 2014, 53(18), 7538-7547.

[5] Jiang, R. Optimization of alarm threshold and sequential inspection scheme. Reliability Engineering & System Safety 2010, 95(3), 208-215.

[6] Rheineck-Leyssius, A. R.; Kalkman, C. J. Advanced pulse oximeter signal processing technology compared to simple averaging. i. effect on frequency of alarms in the operating room. Journal of Clinical Anesthesia 1999, 11(3), 192-195.