Ensembled Self-Adaptive Fuzzy Calibration Models for On-line Cloud Point Prediction (original) (raw)
Related papers
Chemometrics and Intelligent Laboratory Systems, 2013
In melamine resin production process, it is essential to supervise the condensation process. Monitoring the value of the cloud point indicates the best point of time to stop the condensation. Currently, the supervision is conducted manually by operators, which from time to time need to draw and analyze samples from the production process. In order to increase efficiency and to improve quantification quality, in this paper we investigate the usage of non-linear chemometric models, which are calibrated based on near infrared (FTNIR) process spectrum measurements. They rely on fuzzy systems model architecture and are able to incrementally adapt themselves during the on-line process, resolving dynamic process changes which may appear on-line over time due to long-term fluctuations (e.g., caused by dirt) and changes in the composition of the educt, often leading to severe error drifts of static models. Extracting the most informative wavebands prior to model training is essential to avoid a curse of dimensionality; this is achieved by a new extended variant of forward selection, termed as forward selection with bands (FSB). Furthermore, variants of how to integrate auxiliary sensor information (temperature, pH value, pressure) together with the FTNIR spectra are presented (hybridity). A specific ensemble strategy is developed which is able to properly compensate noise in repeated spectrum measurements. Results on high-dimensional data from four independent types of melamine resin show that 1) our non-linear modeling methodology can outperform state-of-the-art linear and non-linear chemometric modeling methods in terms of validation error, 2) the ensemble strategy is able to improve the performance of models without ensembling significantly and 3) incremental model updates are necessary in order to keep the predictive quality of the models high by preventing drifting residuals.
Analytica Chimica Acta, 2018
The physico-chemical properties of Melamine Formaldehyde (MF) based thermosets are largely influenced by the degree of polymerization (DP) in the underlying resin. On-line supervision of the turbidity point by means of vibrational spectroscopy has recently emerged as a promising technique to monitor the DP of MF resins. However, spectroscopic determination of the DP relies on chemometric models, which are usually sensitive to drifts caused by instrumental and/or sampleassociated changes occurring over time. In order to detect the time point when drifts start causing prediction bias, we here explore a universal drift detector based on a faded version of the Page-Hinkley (PH) statistic, which we test in three data streams from an industrial MF resin production process. We employ committee disagreement (CD), computed as the variance of model predictions from an ensemble of partial least squares (PLS) models, as a measure for sample-wise prediction uncertainty and use the PH statistic to detect changes in this quantity. We further explore supervised and unsupervised strategies for (semi-)automatic model adaptation upon detection of a drift. For the former, manual reference measurements are requested whenever statistical thresholds on Hotelling's T 2 and/or Q-Residuals are violated. Models are subsequently re-calibrated using weighted partial least squares in order to increase the influence of newer samples, which increases the flexibility when adapting to new (drifted) states. Unsupervised model adaptation is carried out exploiting the dual antecedent-consequent structure of a recently developed fuzzy systems variant
Improved Calibration of Near-Infrared Spectra by Using Ensembles of Neural Network Models
IEEE Sensors Journal, 2000
Infrared (IR) or near-infrared (NIR) spectroscopy is a method used to identify a compound or to analyze the composition of a material. Calibration of NIR spectra refers to the use of the spectra as multivariate descriptors to predict concentrations of the constituents. To build a calibration model, state-of-the-art software predominantly uses linear regression techniques. For nonlinear calibration problems, neural networkbased models have proved to be an interesting alternative. In this paper, we propose a novel extension of the conventional neural network-based approach, the use of an ensemble of neural network models. The individual neural networks are obtained by resampling the available training data with bootstrapping or cross-validation techniques. The results obtained for a realistic calibration example show that the ensemble-based approach produces a significantly more accurate and robust calibration model than conventional regression methods.
Chemometrics and Intelligent Laboratory Systems, 2011
In polyetheracrylat (PEA) production, it is important to monitor three process parameters in order to assure a high quality of the final product: hydroxyl (OH) number, viscosity and acidity (acid number). Due to the high resolution and high sensitivity, it has been shown in the past that the Fourier transform near infrared (FTNIR) process spectrum measurements can be used to obtain spectra with precise content information about these process parameters. In order to perform an automatic supervision and to reduce the (off-line, laboratory) analysis effort of experts and operators of these substances, chemometric quantification models have to be used. In this paper, we investigate the usage of a specific type of fuzzy systems, so-called Takagi-Sugeno fuzzy systems, for calibrating the chemometric models. This type of model architecture supports the usage of piecewise local linear predictors, being able to model flexibly different degrees of non-linearities implicitly contained in the mapping between NIR spectra and reference values. The training of these models is conducted by an evolving clustering method (adding new local linear models on demand) and a local (weighted) least squares estimation of the consequent parameters, and connected with a wavelength (dimensionality) reduction mechanism. Results on a concrete data set show that it can outperform state-ofthe-art calibration methods as well as support vector regression as alternative non-linear model.
Proceedings IMCS 2012, 2012
Within the industrial research project "Process Analytical Chemistry" (PAC) we are working on FTNIRspectroscopic measurement systems predicting characteristic parameters of industrial production processes. Those parameters are usually monitored offline or at-line with time consuming and expensive laboratory methods. In this contribution, we present a spectroscopic measurement configuration together with the required chemometric analysis, acting as an online-monitoring system. In order to demonstrate the potential of such a system we use the example of melamine resin production in an industrial process. At company partner Dynea the predicted value of the turbidity point is used as an indicator for the end of the batch reaction (turning off heating). Furthermore, we illustrate a way to verify the chemometric prediction by calculating a confidence interval for each predicted value.
Chemometrics and Intelligent Laboratory Systems, 2007
Six popular approaches of «NIR spectrum-property» calibration model building are compared in this work on the basis of a gasoline spectral data. These approaches are: multiple linear regression (MLR), principal component regression (PCR), linear partial least squares regression (PLS), polynomial partial least squares regression (Poly-PLS), spline partial least squares regression (Spline-PLS) and artificial neural networks (ANN). The best preprocessing technique is found for each method. Optimal calibration parameters (number of principal components, ANN structure, etc.) are also found. Accuracy, computational complexity and application simplicity of different methods are compared on an example of prediction of six important gasoline properties (density and fractional composition). Errors of calibration using different approaches are found. An advantage of neural network approach to solution of «NIR spectrum-gasoline property» problem is illustrated. An effective model for gasoline properties prediction based on NIR data is built.
A B S T R A C T The measurements of Near-Infrared (NIR) Spectroscopy, combined with data analysis techniques, are widely used for quality control in food production processes. This paper presents a methodology to optimize the calibration models of NIR spectra in four different stages in a sugar factory. The models were designed for quality monitoring, particularly °Brix and Sucrose, both common parameters in the sugar industry. A three stage optimization methodology, including pre-processing selection, feature selection and support vector machines regression metaparameters tuning, were applied to the spectral data divided by repeated cross-validation. Global models were optimized while endeavoring to ensure they are able to estimate both quality parameters with a single calibration, for the four steps of the process. The proposed models improve the prediction for the test set (unseen data) compared to previously published models, resulting in a more accurate quality assessment of the intermediate products of the process in the sugar industry.
TrAC Trends in Analytical Chemistry, 2002
Near-infrared (NIR) spectroscopy in conjunction with chemometric techniques allows on-line monitoring in real time, which can be of considerable use in industry. If it is to be correctly used in industrial applications, generally some basic considerations need to be taken into account, although this does not always apply. This study discusses some of the considerations that would help evaluate the possibility of applying multivariate calibration in combination with NIR to properties of industrial interest. Examples of these considerations are whether there is a relation between the NIR spectrum and the property of interest, what the calibration constraints are and how a sample-specific error of prediction can be quantified. Various strategies for maintaining a multivariate model after it has been installed are also presented and discussed. #
Strategy for introducing NIR spectroscopy and multivariate calibration techniques in industry
TrAC Trends in Analytical Chemistry, 2003
Analytical methods based on near infrared spectroscopy (NIR) and multivariate calibration techniques have important characteristics for industrial applications because they are quick and can be used to make direct determinations and extract information from several parameters from one single measurement. They can also be used in on-line applications. However, they have not been used in industry as much as one might expect because the validity of the calibration model must be assured when the signal is transformed. There are chemometric techniques for tackling every problem, but there are no specific guidelines about which procedure to use. In this article, we describe a practical strategy for introducing these methods. We discuss the relative merits of the various univariate and multivariate control techniques and describe how they might be used. We also discuss the standardisation of the calibration models and describe how to obtain robust models at low experimental cost. Finally, we illustrate our strategy with a practical case study from the petrochemical industry.
Journal of pharmaceutical and biomedical analysis, 2015
The pharmaceutical industry is under stringent regulations on quality control of their products because is critical for both, productive process and consumer safety. According to the framework of "process analytical technology" (PAT), a complete understanding of the process and a stepwise monitoring of manufacturing are required. Near infrared spectroscopy (NIRS) combined with chemometrics have lately performed efficient, useful and robust for pharmaceutical analysis. One crucial step in developing effective NIRS-based methodologies is selecting an appropriate calibration set to construct models affording accurate predictions. In this work, we developed calibration models for a pharmaceutical formulation during its three manufacturing stages: blending, compaction and coating. A novel methodology is proposed for selecting the calibration set -"process spectrum"-, into which physical changes in the samples at each stage are algebraically incorporated. Also, we esta...