Statistical process fault isolation using robust nonnegative garrote
Introduction
In the last few decades, multivariate statistical process monitoring (MSPM) in modern industrial processes has garnered increased attention due to the requirements for plant safety, product quality, and economics [1], [2], [3]. As discussed in [4], process monitoring consists of four aspects: fault detection, isolation (identification), root cause diagnosis, and process recovery. Fault detection investigates if a fault has occurred in the process. Fault isolation locates the detected fault and identifies the process variable critical to the fault. Root cause diagnosis aims to determine the root cause of the observed process abnormality by using causality analysis techniques, such as Granger causality test, transfer entropy, and Bayesian networks [5], [6], [7], [8], [9]. Then, the effect of the fault can be removed by process recovery. It is obvious that fault isolation is an important step in connecting fault detection to root cause diagnosis, which is the focus of this research work.
There have been a number of classical fault isolation methods. The Mason, Young and Tracy (MYT) decomposition can be applied to fault isolation by dividing the T2 statistic into a limited number of orthogonal items [10]. Despite of its good statistical properties, its practical use in industrial processes can be difficult, because of the huge number of decompositions to be considered [11]. Comparing to the MYT decomposition, contribution plots [12,13] and reconstruction analysis [14,15] are more popular in industrial applications. Contribution plots method isolates the faulty variables by comparing the contribution of each process variable to the monitoring statistic with a predetermined control limit. However, a phenomenon called smearing effect, which means that the faulty variables often influence the contributions of non-faulty variables, limits the effectiveness of this method [16,17]. The cause of smearing effect has been expounded upon in the literature [18]. Reconstruction analysis method can avoid smearing, which is based on the assumption that the candidate fault directions can be acquired. In this method, the faulty variables are isolated by minimizing the reconstructed monitoring statistic along these candidate directions. To relax the requirement of the candidate fault directions, the branch and bound (B&B) algorithm is utilized to locate the faulty variables in reconstruction analysis [19], [20], [21]. However, the computational burden of this algorithm is quite high, especially when there are many variables involved. Therefore, this method is not suitable for online applications. To solve this, the penalized reconstruction-based multivariate contribution analysis method was proposed to improve the efficiency of the fault isolation algorithm [22], [23], [24].
Most recently, Kuang et al. [25] and Yan et al. [26] revealed the equivalence between the problems of multivariate fault isolation and variable selection in discriminant analysis. They illustrated their idea with algorithms based on least absolute shrinkage and selection operator (LASSO). These methods avoid smearing effect and is computationally efficient for online applications, eliminating the need for indicating historical failure data or candidate failures. However, there is no guarantee regarding robustness when the historical normal operating data are contaminated by outliers. Other advanced statistical fault isolation methods, including reconstruction-based contribution [27], modified contribution plot [28], parsimonious reconstruction [29], etc., were not designed for robust fault isolation either.
The problem of robustness has been intensively studied in the field of fault detection [30], [31], [32], [33] but seldom considered in fault isolation. It is commonly assumed that the outliers are identified and removed during the fault detection step before conducting fault isolation. However, sometimes outliers can be carriers of valuable information. Hence, eliminating them could cause information loss during fault isolation. Inspired by the study of [25], this study proposes a nonnegative garrote (NNG)-based fault isolation method, which is further revised to a more robust version using a robust NNG (R-NNG) algorithm. A comparative performance analysis of the LASSO-based fault isolation methods, NNG, and R-NNG is performed using the Tennessee Eastman (TE) process simulation. The results indicate that the proposed R-NNG based fault isolation method outperforms the other two methods.
The rest of this paper is organized as follows. In Section 2, the relationship between fault isolation and variable selection in discriminant analysis is discussed. In Section 3, the NNG and R-NNG-based fault isolation algorithms are proposed. In Section 4, two simulation case studies are used to compare the isolation performance of LASSO, NNG, and R-NNG. Section 5 provides more discussions and concludes this paper.
Section snippets
Problem transformation
According to [25], the multivariate fault isolation problem can be transformed into a variable selection problem in discriminant analysis. The basic idea is to consider two datasets with the same variables—one containing normal operating data and the other containing the detected fault. Fault isolation then identifies the most critical variable that distinguishes the two datasets, which is similar to the choice of the variables in discriminant analysis.
Owing to the relationship between
Penalized regression for multivariate fault isolation
In regression analysis, variable selection is fundamental in making the model more parsimonious and enhancing predictability. Stepwise selection procedures are typically utilized to select a reasonable subset of predictors. Although effective in many cases, stepwise selection has several drawbacks: difficult-to-understand theoretical properties, extremely variable estimates, and high computational costs when there are a large number of predictors [36]. In recent years, penalization techniques
Case study
In this section, the benchmark TE process [42] is used to compare the effectiveness of the presented methods. The diagram of this process is shown in Fig. 1. As a challenging problem for a variety of process control technology studies, the TE process is widely used to test various types of MSPM methods. In this process, two products (G, H) are produced from four reactants (A, C, D, E) with one byproduct (F) being produced simultaneously. Further, a small amount of an inert component (B) is
Conclusions
Fault isolation is an important step in multivariate statistical process monitoring, which identifies the crucial variables responsible for the detected fault. In recent years, variable selection algorithms have been introduced to solve the problems of smearing and lack of historical fault data. The applications of these techniques are promising. However, the robustness issue is seldom addressed. When there are outliers in the historical normal process data, the isolation results can be
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgment
Yao was supported in part by Ministry of Science and Technology, ROC under Grant No. MOST 108-2221-E-007-068-MY3 and the Open Research Project of the State Key Laboratory of Industrial Control Technology, Zhejiang University, China (No. ICT1900319). Wang was supported by Special Project on Industrial Transformation and Upgrading (Made in China 2025) in 2017 (No. TC17085JH). Yang was supported by Innovation Project of Shanghai Science and Technology Committee (No. 18411952200) and Key Research &
References (47)
Survey on data-driven industrial process monitoring and diagnosis
Annu Rev Control
(2012)- et al.
A survey on multistage/multiphase statistical modeling methods for batch processes
Annu Rev Control
(2009) - et al.
Fault diagnosis based on Fisher discriminant analysis and support vector machines
Comput Chem Eng
(2004) - et al.
Advances and new directions in plant-wide disturbance detection and diagnosis
Control Eng Pract
(2007) - et al.
Root cause diagnosis of plant-wide oscillations using granger causality
J Process Control
(2014) - et al.
Generalized contribution plots in multivariate statistical process monitoring
Chemometr Intell Laborat Syst
(2000) - et al.
Analysis of smearing-out in contribution plot based fault isolation for statistical process control
Chem Eng Sci
(2013) - et al.
Bayesian filtering of the smearing effect: fault isolation in chemical process monitoring
J Process Control
(2014) - et al.
A branch and bound method for isolation of faulty variables through missing variable analysis
J Process Control
(2010) - et al.
Reconstruction-based multivariate contribution analysis for fault isolation: a branch and bound approach
J Process Control
(2012)
Variable selection method for fault isolation using least absolute shrinkage and selection operator (LASSO)
Chemometr Intell Laborat Syst
Multivariate fault isolation via variable selection in discriminant analysis
J Process Control
Multivariate fault isolation of batch processes via variable selection in partial least squares discriminant analysis
ISA Trans
Reconstruction-based contribution for process monitoring
Automatica
Fault diagnosis using contribution plots without smearing effect on non-faulty variables
J Process Control
Efficient faulty variable selection and parsimonious reconstruction modelling for fault isolation
J Process Control
Robust probabilistic PCA with missing data and contribution analysis for outlier detection
Comput Stat Data Anal
Variable selection via combined penalization for high-dimensional data analysis
Comput Stat Data Anal
Soft-sensor development with adaptive variable selection using nonnegative garrote
Control Eng Pract
A two-tier approach to the data-driven modeling on thermal efficiency of a BFG/coal co-firing boiler
Fuel
Robust nonnegative garrote variable selection in linear regression
Comput Stat Data Anal
A plant-wide industrial process control problem
Comput Chem Eng
Ensemble deep kernel learning with application to quality prediction in industrial polymerization processes
Chemometr Intell Laborat Syst
Cited by (10)
A hierarchical granger causality analysis framework based on information of redundancy for root cause diagnosis of process disturbances
2024, Computers and Chemical EngineeringData-driven root cause diagnosis of process disturbances by exploring causality change among variables
2023, Journal of Process ControlRoot cause diagnosis of plant-wide oscillations based on fuzzy kernel multivariate Granger causality
2023, Journal of the Taiwan Institute of Chemical EngineersExplainable root cause and pathway analysis with robust and adaptive statistics
2023, Computers in IndustryCitation Excerpt :This task is a real challenge, still under development, and calls for further research and improvements. Several SPC-based root cause identification methods like the Mason, Young and Tracy (MYT), multivariate contribution plots, reconstruction analysis, non-negative garrote-based have been proposed (Wu et al., 2021; Wang et al., 2020; Yan et al., 2018, 2017; Kim et al., 2016; Qiu, 2013). MYT decomposition (Mason et al., 1995), unlike most of these approaches, holds solid statistical properties and foundations.
A fault model extension for a geometric fault isolation methodology to detect leakages and sensor faults on engine test beds
2022, Control Engineering PracticeCitation Excerpt :One solution is to find the variable that distinguishes between the error-free and the faulty data set, similar to the selection of variables in discriminant analysis. Due to the additional relationship between discriminant analysis and regression analysis, a non-negative garrote-based method for fault isolation is proposed that ranks process variables according to how critical they are to the detected fault (Wang et al., 2020). Another solution is to eliminate each variable backwards (Stork, Veltkamp, & Kowalski, 1997) or to treat it as if it were missing (Liu, Chen, & Yao, 2014) and re-estimate the monitoring statistic.
Unlocked decision making based on causal connections strength
2021, European Journal of ControlCitation Excerpt :When a fault is detected, the next step is to look for the variables explaining faulty data. A couple of methods have been proposed for identification including the Mason, Young, and Tracy (MYT) approach [30], multivariate contribution plots [38], reconstruction analysis [19,41,45], and non-negative garrote [36]. Causality analysis techniques were also proposed to identify the variables such as Granger causality test [11], transfer entropy [29], and causal networks [1,25].