Bayesian approach for control loop diagnosis

Abstract

The large number of control loops in a modern industrial plant poses a serious challenge for operators and engineers to monitor these loops to maintain them at optimal conditions continuously. Much research has been done on control loop performance assessment and monitoring of individual components within a control loop. The literature, however, has been sparse in presenting a systematic approach for control loop diagnosis. This thesis is concerned with establishing a data-driven Bayesian approach for control loop diagnosis. Observations from various monitoring algorithms and a priori knowledge of the control loop are synthesized under the Bayesian framework to pinpoint the underlying source of poor control performance. Several challenging practical issues under the proposed framework will also be discussed. To address the incomplete evidence problem that is often encountered in reality, the missing pattern concept is introduced. The incomplete evidence problems are categorized into single missing pattern ones and multiple missing pattern ones. A novel method based on marginalization over an underlying complete evidence matrix (UCEM) is proposed to include the incomplete evidences into the diagnostic framework, such that information in all the evidence samples can be effectively utilized in the diagnosis. Data auto-correlation is common in engineering applications. The temporal information hidden in the historical data is extracted by considering evidence and mode dependency in this thesis. Data-driven algorithms for evidence and mode transition probability estimation are developed. An auto-regressive hidden Markov model is built to consider both mode and evidence dependencies. When both the mode and evidence transitions are considered, the temporal information is effectively synthesized under the Bayesian framework. An approach to estimate the distributions of monitor readings with sparse historical samples is proposed to alleviate the intensive requirement of historical data. The statistical distribution functions for several monitoring algorithm outputs are analytically derived. A bootstrap based method is proposed to handle the challenging problem of estimating the statistical distribution for valve stiction monitoring. The proposed approach has the potential to estimate evidence distribution with as few as only one evidence sample.