Zhengbing Yan

Multivariate fault isolation of batch processes via variable selection in partial least squares discriminant analysis

In recent years, multivariate statistical monitoring of batch processes has become a popular research topic, wherein multivariate fault isolation is an important step aiming at the identification of the faulty variables contributing most to the detected process abnormality. Although contribution plots have been commonly used in statistical fault isolation, such methods suffer from the smearing effect between correlated variables. In particular, in batch process monitoring, the high autocorrelations and cross-correlations that exist in variable trajectories make the smearing effect unavoidable. To address such a problem, a variable selection-based fault isolation method is proposed in this research, which transforms the fault isolation problem into a variable selection problem in partial least squares discriminant analysis and solves it by calculating a sparse partial least squares model. As different from the traditional methods, the proposed method emphasizes the relative importance of each process variable. Such information may help process engineers in conducting root-cause diagnosis.

Valve stiction detection using the bootstrap Hammerstein system identification

Stiction in control valve is one of the long-standing and common problems in the process industries, which accelerates equipment wear and even affects the stability of closed-loop systems. A valve stiction model is proposed to describe the dynamic feature of sticky valve. To detect and quantify valve stiction, a bootstrap Hammerstein system identification procedure is proposed. Through the identified set of the model parameters and operation plant data, the parameter confidence intervals can be predefined and the valve stiction can be easily detected. Numerical examples are provided to illustrate the effectiveness of the proposed method.

A LASSO-based batch process modeling and end-product quality prediction method

Abstract A least absolute shrinkage and selection operator (LASSO) based batch process modeling and end-product quality prediction method is developed in this paper, which overcomes the shortcomings of both multiway partial least squares (MPLS) and the phase-based methods. The proposed phase-LASSO approach models not only the phase characteristics but also the within-phase and between-phase time-dependent information. In the meantime, it automatically selects the critical-to-quality phases via LASSO-type regularization. As a result, phase-LASSO has good prediction accuracy and model interpretation. The effectiveness of the proposed method is illustrated by the case study on injection molding.

Process Fault Isolation via Bayesian Lasso-based Reconstruction Analysis

Abstract In multivariate statistical process monitoring, fault isolation is an important step that identifies the process variables critical to the detected abnormality. Conventionally, contribution plots are the most popular tools for fault isolation, but they often suffer from the smearing effect and give misleading results. Reconstruction analysis is another commonly used technique. Despite its effectiveness, the original reconstruction analysis method relies on the impractical requirement of a sufficient amount of historical fault data or the complete information of candidate fault directions. Recently, the reconstruction analysis technique has been integrated with the least absolute shrinkage and selection operator (Lasso) to overcome its inherent shortcoming. In that research, the task of reconstruction analysis is reformulated as a Lasso problem, and the faulty variables are indicated by the nonzero point estimates of the Lasso coefficients. As well known, a point estimate does not provide any information about its accuracy and is likely to be affected by the quality of the collected data. In this paper, a Bayesian Lasso approach is utilized to solve the problem mentioned above, which assigns independent Laplace (a.k.a. double exponential) priors to the Lasso coefficients and derives the interval estimates by Gibbs sampling from the Bayesian posterior distribution. Such interval estimates can guide fault isolation by conducting statistical hypothesis tests. In addition, the Bayesian framework facilities the tracking of fault propagation path, thereby benefiting the subsequent root-cause diagnosis step.

Robust Fault Isolation using Stable Principal Component Pursuit

Abstract Fault isolation is an important step in multivariate statistical process monitoring, aiming to discovering the process variables critical to the detected fault. However, there are certain shortcomings limiting the implements of the existing methods. Contribution plots often suffer from the smearing effect, while reconstruction analysis requires known fault directions or a large amount of historical fault data that is often unavailable. The performance of the recent developed variable selection method depends on the availability of a historical dataset collected under normal operating conditions and without outlier. Such a dataset is difficult to acquire in real industries. In this research, a robust matrix recovery method called stable principal component pursuit (SPCP) is utilized to solve such problems, which decomposes the data matrix containing both historical operating data and fault measurements into three parts: low-rank process characteristics, sparse errors, and dense noise. In doing so, the variables contributing most to the fault can be identified according to the estimated sparse matrix. The isolation results are robust to the existence of outliers contained in the historical dataset.

Multivariate fault isolation using lasso-based penalized discriminant analysis

Abstract In multivariate statistical process monitoring (MSPC), isolation of faulty variables is a critical step to discover the source of the detected fault. Although fault detection methods have been intensively investigated, studies on fault isolation are relatively limited, due to the difficulty in analyzing the influences of multiple variables on monitoring statistics. To solve the problems of the existing methods, this paper proposes to conduct fault isolation via a lasso-based penalized discriminant analysis technique. Instead of just offering a suggested set of faulty variables, the proposed method provides more information on the relevance of process variables to the detected fault, which facilitates the subsequent root cause diagnosis step after isolation. The benchmark Tennessee Eastman (TE) process is used as a case study to illustrate the effectiveness of the proposed method.