Alarm Event Prediction From Historical Alarm Flood Sequences Based on Bayesian Estimators

Abstract

A new method is proposed to predict upcoming alarm events for a current alarm sequence by exploiting similar historical alarm flood sequences. The proposed method is composed of two main steps. The first step sorts the historical alarm flood sequences in the descending order of similarity scores with respect to the current alarm sequence. The second step chooses some subsequence of recent alarm events in the current alarm sequence and predicts an upcoming alarm event based on these similar historical alarm flood sequences having the same subsequence. A Bayesian estimator is exploited to calculate the posterior probabilities and their confidence intervals for all the candidates of predicted alarm events. The optimal predicted alarm event is the one with the largest lower bound of posterior probability as a quantitative measurement of prediction reliabilities. The proposed method resolves a dilemma between a long (short) subsequence of the current alarm sequence and a small (large) number of the same subsequences in the historical alarm flood sequences. Numerical examples are provided to support the proposed method. Note to Practitioners-Alarm floods are referred to as the phenomena that there are too many alarm events occurring in a short time period such that industrial plant operators cannot promptly handle the overwhelming alarm events. An alarm flood usually means the occurrence of some abnormalities causing negative effects on industrial plants. Hence, it is desirable to predict the upcoming alarm event when several alarm events occur, in order to take proactive actions to prevent a further deterioration or propagation of occurring abnormalities. This article proposes a new method to predict upcoming alarm events for a current alarm sequence based on the Bayesian estimators, by exploiting similar historical alarm flood sequences in the prediction.