Process Alarm Event Prognosis Based On Bayesian Network

With the increasing complexity of modern industrial processes, operational safety should be paid more and more attentions. As main tools of guaranteeing the safety of industrial processes, reliable alarm systems play an important role in improving process operational performances. Specifically, alarm predictions can not only detect dangerous situations earlier but also guide human operators to take appropriate precautions, thereby effectively reducing the probability of accidents.In this thesis, an approach to predicting alarm events based on Bayesian network models is explicitly introduced. Firstly, Bayesian network learning algorithms are investigated and implemented, illustrating the expected reasoning performances of Bayesian networks. Subsequently we extract sequences of alarm events from historical process data, which help establish Bayesian networks corresponding to single-variable and multi-variable alarm events, respectively. Parameters and structures of the Bayesian networks are determined by combining expectation-maximization (EM) algorithms and greedy search algorithms. Taking advantage of the networks, probability reasoning procedures are performed to predict process alarm events. Case studies on an industrial DMF recovery process and the TE process are carried out, employing the proposed methods in predicting industrial alarm events. Simulation results show that the proposed methods can help extract useful information from industrial data so as to predict alarm events with high accuracy.