LAUSR

Abstract In this paper, a novel robust Bayesian network is proposed for process modeling with low-quality data. Since unreliable data can cause model parameters to deviate from the real distributions and make network structures unable to characterize the true causalities, data quality feature is utilized to improve the process modeling and monitoring performance. With a predetermined trustworthy center, the data quality measurement results can be evaluated through an exponential function with Mahalanobis distances. The conventional Bayesian network learning algorithms including structure learning and parameter learning are modified by the quality feature in a weighting form, intending to extract useful information and make a reasonable model. The effectiveness of the proposed method is demonstrated through TE benchmark process and a real industrial process.