Eliminating the Effect of Multivariate Outliers in PLS-Based Models for Inferring Process Quality
TL;DR: In this paper, the authors proposed a robust modeling method based on PLS, which not only alleviates the harmful effect of multivariate outliers, but also retains the information necessary for building a robust model from the training data.
Abstract: Outliers in multivariate data demand special attention in data-driven process modeling Their extremeness usually gives them an excessively high influence in the calculation, which may result in a less precise model It is challenging to detect them using existing univariate approaches A novel robust modeling method is presented; this PLS based modeling procedure not only alleviates the harmful effect of multivariate outliers, but also retains the information necessary for building a robust model from the training data The performance of the proposed approach is compared with conventional strategies using an actual industrial case study
Citations
More filters
TL;DR: A novel soft sensor technology based on partial least squares (PLS) regression is developed and applied to a refining process for quality prediction and the performance of the resulting soft sensor is evaluated by comparison with laboratory data and analyzer measurements.
Abstract: In the petrochemical industry, the product quality reflects the commercial and operational performance of a manufacturing process. However, real-time measurement of product quality is generally difficult. Online prediction of quality using readily available, frequent process measurements would be beneficial in terms of operation and quality control. In this paper, a novel soft sensor technology based on partial least squares (PLS) regression is developed and applied to a refining process for quality prediction. The modeling process is described, with emphasis on data preprocessing, multivariate-outlier detection and variables selection. Enhancement of PLS strategy is also discussed for taking into account the dynamics in the process data. The proposed approach is applied to data from a refining process and the performance of the resulting soft sensor is evaluated by comparison with laboratory data and analyzer measurements.
158 citations
Cites methods from "Eliminating the Effect of Multivari..."
...The PLS regression technique explains variations in both and simultaneously and also maximizes the and covariance....
[...]
TL;DR: The proposed approach uses samples from a given state to develop a description of the non-self-space in the form of a collection of spherical detectors, in contrast to traditional statistical and pattern recognition algorithms that store descriptions of the space occupied by the normal samples.
Abstract: Artificial immune system is a new artificial intelligence methodology that is increasingly attracting much attention for monitoring engineered systems. In an artificial immune system (AIS), principles and processes of the natural immune system are abstracted and applied in solving real world problems. One immune-inspired principle is negative selection, whereby the natural immune system distinguishes between the body's own (self) cells and foreign (non-self) cells. In this paper, we apply this principle for process monitoring and fault diagnosis. In the proposed approach, samples from a given state (such as normal or known fault) are considered as self. The proposed approach uses these samples to develop a description of the non-self-space in the form of a collection of spherical detectors. This representation is in contrast to traditional statistical and pattern recognition algorithms that store descriptions of the space occupied by the normal samples. The proposed fault detection and identification (FDD approach is a generic one and can be applied for monitoring and fault diagnosis of both continuous as well as batch processes and transient operations since it does not require that the underlying data stems originate from a specified statistical distribution. The effectiveness of the proposed approach for monitoring and fault diagnosis is demonstrated through various case studies. The results of the case studies clearly illustrate the method's ability to provide excellent monitoring and diagnosis performances with (i) complete fault coverage (all the faults studied can be readily detected and identified), (ii) very high overall recognition rate, (ii) low false positive rate, (iii) high true positive rate. and (iv) early fault detection and diagnosis. A comparison of performance with traditional principal component analysis (PCA) based approaches is also performed.
28 citations
TL;DR: PLSR process model is chosen as the basis of the difficult-to-measure process variable estimator while its parameters are updated in several ways—by the moving window method, recursive NIPALS algorithm, recursive kernel algorithm and Just-in-Time learning algorithm.
Abstract: Problemi s upravljanjem mnogih procesa u industriji vezani su s nemogucnoscu on-line mjerenja nekih važnih procesnih velicina. Ovaj se problem može u znacajnoj mjeri rijesiti estimacijom ovih tesko-mjerljivih procesnih velicina. Estimator je pri tome odgovarajuci matematicki model procesa koji na temelju informacije o ostalim (lako-mjerljivim) procesnim velicinama procjenjuje trenutni iznos tesko-mjerljive velicine. Buduci da su procesi po prirodi promjenjivi, tocnost estimacije zasnovane na modelu procesa izgrađenog na starim podacima u pravilu opada s vremenom. Kako bi se ovo izbjeglo, parametre modela procesa je potrebno kontinuirano prepodesavati kako bi model sto bolje opisivao (trenutno) vladanje procesa. Ovisno o tipu matematickog modela, za prepodesavanje njegovih parametara na raspolaganju je vise metoda. Kao osnova estimatora tesko-mjerljive velicine u radu se koristi PLSR model procesa, dok se njegovi parametri prepodesavaju na vise nacina – metodom pomicnog prozora, rekurzivnim NIPALS algoritmom, rekurzivnim kernel algoritmom te Just-in-Time Learning metodom. Svojstva navedenih metoda adaptacije PLSR modela procesa ispitana su na odabranom primjeru. Nadalje, metode adaptacije su analizirane i s obzirom na racunalnu i memorijsku zahtjevnost.
5 citations
Dissertation•
01 Jan 2014
TL;DR: In this paper, the authors present a table of acknowledgments for the authors of this paper: https://www.sal.org.au/ acknowledgments/Acknowledgments.
Abstract: ...................................................................................... i Acknowledgments.......................................................................... ii Table of
3 citations
TL;DR: In this work, process dynamics was also identified in latent subspace using neural networks and the inverse dynamics of the latent variable based NN process acted as inverse neural controller (DINN).
Abstract: Partial least squares technique has been in use for identification of the dynamics & control for multivariable distillation process. Discrete input-output time series data ) ( Y X were generated by exciting non-linear process models with pseudo random binary signals. Signal to noise ratio was set to 10 by adding white noise to the data. The ARX models as well FIR models in combination with least squares technique were used to build up dynamic inner relations among the scores of the time series data ) ( Y X , which logically built up the framework for PLS based process controllers. In this work, process dynamics was also identified in latent subspace using neural networks. The inverse dynamics of the latent variable based NN process acted as inverse neural controller (DINN). Distillation process without any decoupler could be controlled by a series of NN-SISO controllers General Terms Process Identification & control, Statistical Process Control
3 citations
Cites methods from "Eliminating the Effect of Multivari..."
...Wang & Srinivasan (2009) proposed a novel method that provides a robust model and retains the essential information from the data....
[...]
References
More filters
TL;DR: The robust estimation of the covariance or correlation matrix at each scale is obtained by the proposed approach so that accurate MSPCA models can be obtained for process monitoring purposes.
Abstract: Robust multi-scale principal component analysis (RMSPCA) improves multi-scale principal components analysis (MSPCA) techniques by incorporating the uncertainty of signal noise distributions and eliminating/down-weighting the effects of abnormal data in the training set. The novelty of the approach is to integrate MSPCA with the robustness to the typical normality assumption of noisy data. By using an M-estimator based on the generalized T distribution, RMSPCA adaptively transforms the data in the score space at each scale in order to eliminate/down-weight the effects of the outliers in the original data. The robust estimation of the covariance or correlation matrix at each scale is obtained by the proposed approach so that accurate MSPCA models can be obtained for process monitoring purposes. The performance of the proposed approach in process fault detection is illustrated and compared with that of the conventional MSPCA approach through a pilot-scale setting.
80 citations
01 Aug 2006
TL;DR: In this article, a soft sensor technology based on partial least squares (PLS) regression between process variables and quality variable is developed and applied to a refinery process for quality prediction, with emphasis on data preprocessing, PLS regression, multi-outliers detection and variables selection in regression.
Abstract: In petrochemical industry, the product quality encapsulates the commercial and operational performance of a manufacturing process. Usually, the product quality is measured in the analytical laboratory and it involves resources and considerable time delay. On-line prediction of quality using frequent process measurements would be beneficial in terms of operation and quality control. In this article, a novel soft sensor technology based on partial least squares (PLS) regression between process variables and quality variable is developed and applied to a refinery process for quality prediction. The modeling process is described, with emphasis on data preprocessing, PLS regression, multi-outliers' detection and variables selection in regression. Enhancement of PLS is also discussed to take into account the dynamics in the process data. The proposed approach is applied to data collected from a refinery process and its feasibility and performance are justified by comparison with laboratory data.
11 citations