Scientific communications

In a previous study in 2018, we compared several Chemometrics and Machine Learning methods: Linear PLS modeling, X-variable transformation – PLS to cope with non-linearity, Local Modeling (Locally Weighted Regression – LWR), Random Forest (RF), Support Vector Machines (SVM) and Artificial Neural Networks (ANN) on Near-Infrared (NIR) spectroscopic data. We proved that ML techniques can significantly improve the quantitative model performance for non-linear modeling relationships.

In this study, we will focus particularly on Support Vector Machines (SVM), since they show several advantages: they are able to model non-linear relationships, with rather good performances over linear methods, but require fewer samples and less optimization burden than ANN.

Chemometrics tools are thus used to combine the different data tables to extract the best information to make decisions. Different strategies are possible to explore a set of data tables. The most simple and intuitive idea is to concatenate or merge data and perform a PCA using an appropriate weighting strategy. Other methods have been developed to take into account within the algorithm the fact that data comes from different tables and to estimate the relationship between the tables. For example, Qannari and al. developed the Common Component and Specific Weight Analysis (CCSWA) and Wold et al. worked on the Hierarchical PCA (HPCA). Some authors are trying now to unify the theory and explain the links between all these methods.

The present work comes from the implementation of the workshop called ‘Multi-block analysis for combining different kinds of spectroscopic data tables’ held during the pre-conference courses of the NIR2013 conference in La Grande Motte, France. We tried to show in a didactic way, as a tutorial, the key issues addressed to the end-user for multi-block analysis used in exploratory applications. The different steps and pitfalls are illustrated using a real data table set collected on virgin olive oils coming from different regions.

Model corrections with classical methods such as bias and slope or model redevelopment are not always satisfactory.

The use of an orthogonalization method can be a very effective way to solve this problem; this approach is illustrated in this study with an industrial application

This project was carried out with the French Wine and Vine Institute (IFV). It aims at modeling the qualitative potential of a vineyard plot by relating data from Mid-Infrared spectra measured on grapes before harvest to the final quality of the wine.

The objective was to build a model to discriminate the mid-infrared spectra and a sensory classification with 2 classes (Fruity/ Vegetal) over several vintages.

The problems occurred for this calibration transfer in a real case within the framework of a discrimination model will be detailed. A focus will be especially made on the importance of the choice of the method transfer (DS, PDS, TOP / EPO, local centering, etc.) and the choice of samples for the standardization set.

The data measured with these sensors must be compiled and analyzed, in order to carry out the process supervision in globality. Thus this monitoring can be performed using multivariate methods such as MSPC (Multivariate Statistical Process Control) for continuous production processes or BSPC (Batch Statistical Process Control) for batch processes. These methods optimize and supervise the production processes. They also provide in a diagnosis of drift causes, taking into account all the process parameters.

Ten years after, PAT has been developed, tested and implemented in numerous R&D and industrial sites all over the world.

The first part of this presentation gives details on PAT goals and implementation. The second part focuses on one important field of application of PAT with high added value, i.e. batch monitoring or Batch Statistical Process Control (BSPC) using multivariate data analysis.

Due to the various sources of process variability impacting the process trajectory, none of the blends will reach homogeneity at the same time. Some qualitative methods (Maesschalck, 1998; Sekulic, 1998; Puchert, 2011), e.g. PCA, SIMCA, and quantitative methods (Berntsson, 2002), e.g. PLS, associated with specific statistics have already been applied to help determine when a blending operation should be terminated. This study proposes to use a qualitative method called Prototype using innovative statistics for end-point determination.

During the continuous process control, the difficulty but also the interest of on-line NIR analysis is to follow the transition from a product A to a product B, in order to control the conformity of these products.

For polymers, the transitions may be due to a change in the composition (% of comonomer X relative to Y, addition or deletion of a comonomer). It also can be the result of a change in the polymerization rate. Comprehensive models for the quantitative prediction of polymer properties allow ​​to obtain values throughout the process; however, in most cases, these models are not acurrate enough. It is therefore important to determine which product the NIR analyzer is facing in order to choose the specific quantitative model of the product and thus predict the end of the transition and the transition to a stabilized state.

This article aims to assess various polymers discrimination strategies based on NIR spectra during transitions in order to choose the appropriate quantitative methods for online polymerization monitoring.

This study makes it possible to tackle the problem of discrimination and calibration transfer in mid-infrared spectroscopy in a complex framework, with the following constraints:

• Characterization of an aromatic potential using mid-infrared spectra
• Discrimination model based on a sensory reference measurement
• Calibration transfer between spectrometers (various equipment manufacturers of spectrometers, network of same spectrometers between cooperative cellars and laboratories, management of maintenance operations / breakdowns)
• Availability of standardization samples
• Representativeness of sampling
• Vintage variability

The objective is to establish a discrimination model based on mid-infrared spectra and a sensory classification with 2 classes (Fruity / Vegetal) over several vintages.

The problem of the calibration transfer, within the framework of a discrimination model, is detailed, in particular the importance choosing the transfer method (DS, PDS, TOP / EPO, local centering …) and the standardization sample set.

This work was carried out as part of a collaborative FUI research project called VINNEO, funded by the French state, the Occitanie Region (ex Languedoc-Roussillon), BPI France and the European FEDER Fund, in collaboration with the teams of Vinovalie, a wine cooperative cellars in the South-West.