Dynamic Bayesian Network for Robust Latent Variable Modeling and Fault Classification
Junhua Zheng,Jinlin Zhu,Guangjie Chen,Zhihuan Song,Zhiqiang Ge
DOI: https://doi.org/10.1016/j.engappai.2020.103475
IF: 8
2020-01-01
Engineering Applications of Artificial Intelligence
Abstract:This work deals with robust dynamic probabilistic modeling and fault classification for process data. In dynamic processes, observed variables can be numerous in amount and correlated with each other in both variable-wise and time-wise. While multivariate statistical analysis methods such as principle component analysis (PCA) are widely used to handle variable-wise correlations and also for dimension reduction purpose, the time-wise correlation among process data is typically addressed by constructing dynamical models, e.g. time-series data model, state space model, etc. In this paper, the robust probabilistic principle component analysis (RPPCA) model is introduced for feature extraction from contaminated process data, suffered from outliers, disturbances, heave noises, etc. A dynamic Bayesian network (DBN) is then constructed, with incorporation of a mixture of Gaussian components for approximation of the non-Gaussian characteristics of latent variables. Based on the developed robust dynamic Bayesian network model, a fault classification scheme is proposed. To validate the effectiveness and feasibility of the new method, two well-known benchmark case studies are carried out. Simulation results show that robust dynamic method performs better in outlier contaminated situations, where the performance has been improved by 10%–20% in most cases.