A novel scoring function based on family transfer entropy for Bayesian networks learning and its application to industrial alarm systems. (April 2019)

• Record Type: Journal Article
• Authors: Meng, Qian-Qian; Zhu, Qun-Xiong; Gao, Hui-Hui; He, Yan-Lin; Xu, Yuan
• Is Part Of: Journal of process control. Volume 76(2019)
• Page Start: 122
• Page End: 132
• Publication Date: 2019-04
• Subjects: Bayesian networks -- Scoring function -- Structure learning -- Transfer entropy -- Alarm systems; Process control -- Periodicals; Fabrication -- Contrôle -- Périodiques; Process control; Periodicals; Electronic journals; 660.281
• DOI: 10.1016/j.jprocont.2019.01.013
• Languages: English
• ISSNs: 0959-1524

Abstract

Highlights: Scoring function using transfer entropy for Bayesian networks learning is proposed. A penalty term is added to scoring function to avoid overfitting during learning. Errors brought by the self-interference of alarm variables can be much reduced. Alarm sequences with three states are used to reduce the computational complexity. Simulation results verify the feasibility and effectiveness of the proposed method. Abstract: Bayesian network (BN) is a powerful reasoning and knowledge expression tool combining the graph theory and the probability theory. Establishing an accurate Bayesian network for alarm systems plays a critical part of alarm root cause analyses in industrial processes. Bayesian networks are hard to learn, because current states of alarm variables are influenced not only by other variables but also by the history states of themselves. In order to handle this problem, a novel scoring function named Family Transfer Entropy Tests (FTET) for Bayesian networks learning is proposed. In the proposed FTET scoring function, the family score (FC) of each family in a Bayesian network is defined using Family Transfer Entropy (FTE). FTE is used to quantify the degree of the interaction between variables. Moreover, in the proposed FTET, FTE with penalty is considered to avoid overfitting in Bayesian network learning. In order to validate the performance of the proposed FTET scoring function, case studies based on a stochastic process and the Tennessee Eastman (TE) process are carried out. Simulation results show that the errors brought by the impact of the history states of the variable itself are reduced. The Bayesian network structure learnt from the proposed FTEF scoring function is simpler and more accurate compared with that learnt from the well-known scoring functions of Bayesian Information Criterion (BIC) and Bayesian Dirichlet (BDe).