FR2953929A1 - Spectroscopic sample e.g. food, analyzing method for measuring changes in nutritional and toxicology properties of food, involves determining content of sample, where average spectral difference between light radiations is fifty nanometers - Google Patents

Abstract

The method involves lighting a sample (E) to be analyzed by excitation light radiations, and acquiring front fluorescence spectrum of the sample. The acquired fluorescence spectrum is pre-treated, and a vector of scores by applying to the pre-treated spectrum of a multi-channel statistical model identified by a loading vector of fluorescence excitation and a vector of emission loadings is calculated. Content of the sample in a substance is determined from the score vector, where the average spectral difference between the radiations is 50 nanometers on a spectral range of 100 nanometers. An independent claim is also included for an apparatus for spectroscopic analysis of a sample.

Description

The invention relates to a method and an apparatus for the spectroscopic analysis of at least one sample, in particular of a food or medicament, and a method for the spectroscopic analysis of at least one sample. , implementing a spectroscopic data analysis method based on a multi-channel statistical model. The invention can be applied in particular, but not only, to the study of the evolution of the nutritional properties (vitamin content ...) and / or toxicological (content of neoformed contaminants ...) of a food at during its preparation or conservation. Neoformed contaminants (CNF) are new molecules formed in the food from the clean components of this food, due to the action of transformation processes, and which pose public health problems. Few public or private laboratories are able today to measure these contaminants. In addition, although several methods have been published for the determination of each of the contaminants, none of them has actually been subject to standardization and standardization. In addition, these analyzes are expensive and the time to obtain the results is long (2 to 3 weeks). The industry requires faster and less expensive testing methods to ensure compliance with its processes and the safety of the food it produces. Food processing processes (cooking, sterilization, preservation) can also adversely affect the nutritional properties of foods, for example by lowering their vitamin levels. The characterization of nutritional properties poses problems similar to those mentioned above for the determination of neoformed contaminants. It is known to perform analyzes, especially spectroscopic, of foods using the methods of chemometrics, and in particular to multi-channel analysis. Multichannel analysis is the natural extension of multivariate analysis when data is arranged in tables with three or more channels. It is based on the use of statistical models such as PARAFAC (Parallel Factor) and NPLS (N-ways Partial Least Squares Regression), ie regression partial least squares to n-ways). These methods, as well as their application to the analysis of food products, are described in [Bro 1998].

Specifically, [Rizkallah 2007] describes a method of analyzing unprepared samples (only milled, if necessary) based on the application of the PARAFAC model to frontal fluorescence fluorescence spectra. This method involves illuminating a sample with a plurality of monochromatic light rays at respective wavelengths, and acquiring the corresponding fluorescence spectra. The excitation radiations finely sample (several hundred points) a spectral range covering the visible and the near UV; in turn, the fluorescence emission spectra are spectrally sampled (also over several hundred points).

The spectral data thus collected for each sample are arranged in a large matrix known as the "excitation-emission matrix" (MEE), one dimension of which represents the emission wavelengths and the other dimension represents the wavelengths. resignation. MEAs corresponding to several calibration samples are analyzed by the PARAFAC method which extracts information, called "PARAFAC factors", which correspond to bilinear fluorescence profiles and their relative intensities. Then, a multiple regression between the fluorescence intensities and the concentrations of neoformed contaminants (or other substances of interest) measured chemically makes it possible to construct a calibration model that is then used to predict the concentrations of neoformed contaminants from the collected EEMs. on new samples to analyze. This method has two major limitations: - First, it requires a high cost equipment: a spectrofluorometer equipped with a Xenon lamp delivering all the wavelengths of UV near to visible, and two monochromators capable of scrolling these different wavelengths at excitation as at emission for the photons emitted by the sample. - Second, its implementation is long because, even if sample preparation is avoided, the acquisition of each EEM takes 15 to 45 min depending on the desired spectral resolution. The analysis of these data matrices, including a considerable number of variables, is also very heavy despite the automation of the data processing algorithms. A similar method is described by [Rizkallah et al., 2008]. The document [Nahorniak 2003] studies the influence of the spectral resolution (and thus the size of the matrix EEM) on the error of prediction of the concentration of a fluorophore by application of the PARAFAC method. In this document, diluted solutions of simple and known composition are considered; in addition, all the measurements are made with a high spectral resolution, of 2 nm / pixel, the reduction of resolution being obtained by calculating spectral averages. It follows from this study that the prediction error is minimized and is about 0.5-1% for an excitation resolution of 6 to 10 nm / pixel, and then increases strongly for lower resolutions (one factor increase). 2 to 6 at a resolution of 20 nm / pixel). It therefore seems difficult to use resolutions below 20 nm / pixel. The aim of the invention is to provide a simpler, faster analysis method requiring simpler and less expensive equipment, and therefore better suited to industrial requirements. According to the invention, such a goal is achieved by a method of spectroscopic analysis of at least one sample using a method for analyzing spectroscopic data based on a multi-channel statistical model, the method comprising: a) lighting said or each sample to be analyzed by a plurality of excitation light beams at respective wavelengths; b) acquiring frontal fluorescence spectra of said or each sample, each corresponding to a respective excitation light radiation; c) pretreatment of the acquired fluorescence spectra; d) for each sample, calculating a score vector (or light intensity) by applying to the pretreated spectra of said multi-channel statistical model, identified by a fluorescence excitation loadings vector and a vector of emission loadings (or spectrum); and e) determining the content of said or each sample in at least one substance of interest from said score vector; the method being characterized in that the average spectral difference between said excitation light beams is at least 20 nm (preferably at least 50 nm) over a spectral range of at least 100 nm. Advantageously, the number of excitation light rays and corresponding fluorescence spectra for each sample may be between two and six, and preferably between three and five, over a spectral range of at least 100 nm, and from preferably at least 150 nm. This method is innovative in that it uses multichannel analysis tools such as the PARAFAC model, normally used for large-scale MEEs, to a set consisting of a limited number of spectra (usually two to six, preferably three to three). and five), sampling in a very coarse way the spectral range of excitation of the samples. This range is relatively wide (at least 100 or even 150 nm or more) because the physicochemical composition of said samples is largely unknown, so the exact position of the absorption spectra of the fluorophores contained in said sample is also unknown. . The present inventors have surprisingly found that multichannel models continue to be applicable and provide usable results even using such a coarse excitation spectral resolution, and such a limited number of excitation wavelengths. It is important to emphasize that from a physical point of view, the loadings represent the emission or excitation spectra of each factor, and the "scores" the intensities of these factors. A "factor

PARAFAC "represents the contribution of a fluorophore, or a mixture of fluorophores, common to the analyzed samples. It must also be emphasized that the fluorescence spectra are deformed by the interaction between the excitation photons and the photons emitted with the food matrix and that the generally accepted trilinearity condition for the application of the PARAFAC model is not satisfied. Nevertheless, the inventors have shown that this model is usefully applied to such matrices EEM to describe the fluorophore composition. It should be understood that the process of the invention is not intended (at least not necessarily) to determine the content of the fluorophores themselves. This method simply exploits correlations, generally non-linear or multilinear, between the evolutions of the excitation and fluorescence emission spectra and the content of nutrients and / or contaminants which, most often, are not fluorescent.

A comparison with the aforementioned document [Nahorniak 2003] is particularly interesting. This document relates to a particularly favorable case, namely the analysis of solutions of simple and known composition. On the contrary, the invention relates to the frontal fluorescence analysis of complex and highly diffusing samples, the composition of which is largely unknown. In addition, in the document [Nahorniak 2003], the spectra are acquired at high excitation spectral resolution, then averaged, which is likely to reduce the prediction error. In contrast, in the case of the invention, only two to six spectra are acquired, with an extremely low excitation spectral resolution, less than 20 nm / pixel or even 50 nm / pixel. Despite this, a timely choice of parameters of the multichannel model and / or a timely pretreatment of the data makes it possible to limit the prediction error to quite acceptable values, typically less than 15% or even 10%. A method according to the invention may also include a preliminary calibration phase comprising: i) illuminating a plurality of calibration samples by said excitation light beams; ii) acquisition of the frontal fluorescence spectra of the samples, corresponding to said excitation light rays; iii) pretreatment of the acquired fluorescence spectra; iv) determining, by an iterative method, said excitation and emission loadings vectors of the multipath model, as well as a score vector for each calibration sample; and v) determining a regression function binding said score vectors to known levels of at least one substance of interest in said calibration samples.

Advantageously, the fluorescence spectra can be pretreated by subtracting a contribution due to the first order Rayleigh scattering of the excitation light radiation, said contribution being calculated by means of a generalized linear model. In particular, said generalized linear model can be determined from a first portion of the spectrum to be pretreated, comprising only said contribution due to Rayleigh scattering, and is used to predict the Rayleigh scattering contribution in a second portion of said spectrum in which it is superimposed on a fluorescence contribution. The subtraction of the Rayleigh scattering is also necessary in the known prior art methods of analysis, but becomes critical in the method of the invention. The use of a generalized linear model, replacing known replacement techniques with zeros or missing values, is particularly advantageous when a very small number of excitation wavelengths are used, as in the case of the invention.

Then, the fluorescence spectra can also be pretreated by normalizing them. Several samples, calibration and / or to be analyzed, can also be pretreated by performing a multiplicative scatter correction (MSC). According to a preferred embodiment of the invention, said multi-channel statistical model can be a PARAFAC model, the fluorescence spectra of each sample being represented in concatenated form. In this case, the loadings vectors of said PARAFAC model can be determined by an iterative method which is stopped when the decrease of the loss function due to the last iteration drops below a threshold value ("convergence parameter" ) between 10-4 and 10-2. This value should be compared to that used generally from 10-6. Indeed, the use of a larger convergence parameter facilitates the convergence of the PARAFAC model when a small number of excitation wavelengths is used. Alternatively, said multichannel statistical model may be an NPLS model, the fluorescence spectra of each sample being represented in concatenated form. In this case, the fluorescence spectra can usefully be pretreated by orthogonal correction of the signal. Advantageously, said excitation light beams may comprise: a first radiation at a wavelength of between 270 and 300 nm; a second radiation at a wavelength of between 300 and 360 nm; and a third radiation at a wavelength of between 400 and 500 nm. The samples to be analyzed and calibrated can be, in particular, samples of a product selected from a food and a drug. Another object of the invention is the use of a method as described above for measuring the evolution of the nutritional and / or toxicological properties of a food during its preparation or preservation. Yet another object of the invention is an apparatus for the spectroscopic analysis of at least one sample comprising: a set of light sources for illuminating said or each sample to be analyzed by respective excitation light beams, having lengths of different waveforms with an average spectral difference of at least 20 nm over a spectral range of at least 100 nm; means for acquiring the frontal fluorescence spectra emitted by said or each sample when it is illuminated by said excitation light beams; and a means for processing the fluorescence spectra acquired, adapted to implement a method as described above. Preferably such an apparatus may comprise between two and six (and preferably between three and five) of said light sources.

Advantageously, such an apparatus may comprise: a first light source emitting radiation at a wavelength of between 270 and 300 nm; a second light source emitting radiation at a wavelength of between 300 and 360 nm; and a third light source emitting radiation at a wavelength of between 400 and 500 nm. Other features, details and advantages of the invention will emerge on reading the description made with reference to the accompanying drawings given by way of example and which represent, respectively: FIG. 1, three "raw" fluorescence spectra d a sample of roasted chicory, corresponding to three excitation radiations at different wavelengths; FIGS. 2A, 2B, 2C, three graphs illustrating the subtraction operation of Rayleigh scattering by the method using a generalized linear model; FIGS. 3A, 3B, 3C, three graphs comparing said Rayleigh scattering subtraction method (FIG. 3C) with two other methods known from the prior art: the substitution with zeros (FIG. 3A) and the replacement with missing values (Figure 3B); Figure 4 is a schematic illustration of the notion of concatenation of spectra; FIG. 5, a schematic illustration of the PARAFAC method; - Figure 6, the block diagram of an analysis apparatus 25 according to one embodiment of the invention; FIGS. 7 and 8, graphs illustrating the application of a method according to a first embodiment of the invention, based on a PARAFAC model, to the prediction of the acrylamide content of various chicory samples; FIGS. 9A, 9B, 10A and 10B, graphs illustrating the application of a method according to a second embodiment of the invention, based on an NPLS model, to the prediction of the vitamin C content of different green bean samples.

The method of the invention uses the fluorescence signal emitted on the surface of the food after illumination by light beams at determined wavelengths of the UV-visible range (approximately 250 - 750 nm). This signal is analyzed by chemometric methods that extract the information that is correlated with the quality parameters that are to be measured in the food. The existence of such a correlation is deduced from the fact that, during agricultural production, conservation and transformation, the intrinsic fluorescence of the natural constituents of the food (vitamins, proteins and other natural constituents or intentionally or unintentionally added ), as well as their reflectance, evolve, while at the same time, new signals appear due to the formation of new molecules (including newly formed contaminants). Fluorescence is neoformed or acquired depending on whether fluorophores are de novo or come from the environment. The joint evolution of native (SNF) and newly acquired (SNA) signals is robustly correlated with physico-chemical changes in the food, in particular with changes in quality parameters, induced during production, conservation and / or processing. The factors of change that influence the quality of the food are oxidizing ultraviolet radiation, the destruction of microorganisms or, on the contrary, the development of some of them that can lead to the synthesis of mycotoxins, human intervention on crops. (fertilizers, pesticides ...), or the application of processes modifying the temperature, the pressure or any other physical parameter within the food and which consequently induce a modification of the physico-chemical composition and the parameters quality. The excitation light beams have selected wavelengths so as to explore the UV-visible spectrum as widely as possible. In general, one can choose a priori: • a wavelength between 270 and 300 nm to excite tryptophan, phenols such as chlorogenic acid or hydroxytyrosol or vitamin E, molecules emitting in l UV;

A wavelength of between 400 and 450 or 500 nm for exciting riboflavin, porphyrins and chlorophyll, molecules emitting in the visible range (500-700 nm); One or two wavelengths between 300 and 360 or even 400 nm are introduced to excite the neoformed fluorophores, mainly Maillard products and lipid peroxidation products, but also the mycotoxins emitting in the visible far-near UV ( 400-500 nm). These wavelengths will be modified in case of applications other than the food industry. In any case, for the sake of precision, it is possible to choose the wavelengths as close as possible to the maxima of the loadings representing the excitation vector obtained by the PARAFAC decomposition of a complete MEE matrix obtained with a laboratory fluorometer on a a batch of representative samples.

The number of wavelengths used is between 2 and 6, advantageously between 3 and 5, and preferably equal to 5, in order to excite in turn a corresponding number of groups of fluorophores among those described above. The intensity of the excitation radiation is chosen so that the fluorescence emission energy of these fluorophores is significantly modified during the transformation steps that are to be characterized. For example, the intensity must be sufficient for the appearance of mycotoxins related to the development of fungi on a seed during its transformation substantially modifies the fluorescence emission collected on the detector from an uncontaminated sample.

It must also make it possible to highlight the application of such a heat treatment characterized by a time-temperature pair; or the application of a washing, bleaching or decontamination treatment. Figure 6 shows a very simplified diagram of an apparatus for carrying out the method of the invention. This apparatus comprises three light sources SI, S2, S3 each emitting a monochromatic radiation beam at a different wavelength directed to illuminate the sample E. The fluorescence signal F (in fact, fluorescence mixture and Rayleigh scattering 1st and sometimes 2nd order) issued by this sample is 2953929 It is transported by an optical fiber to a spectrometer which breaks down the light radiation emitted in spectrum. The acquired spectra are processed by a MTD data processing means (typically a computer programmed in a timely manner) which makes it possible to extract the desired chemometric information. The light sources can be, typically, light-emitting diodes, or even lasers (preferably semiconductor) if higher intensities are required. Spectroscopic data analysis, performed by the MTD computer, has four main steps: pretreatment of the spectra; the determination of the loadings vectors of the multipath model using the spectra of the calibration samples; the determination of the vectors of the "scores" by applying this model to the spectra of the samples to be analyzed; and determining the information of a chemical nature (content of one or more substances of interest) from said "score" vectors. The pretreatment of the spectra is first considered, with reference to a specific example, in which a roasted chicory sample is illuminated successively by three excitation radiations at 280, 340 and 429 nm. Each of the three fluorescence spectra consists of 1515 spectral intensity values for as many different wavelengths. The spectral resolution is 0.25 nm / pixel, but it can be divided by 5 or more, without any degradation of the results of the process of the invention being observed. The "raw" spectra (Figure 1) are dominated by first-order Rayleigh scattering at the wavelength of the excitation radiations (280, 340 and 429 nm). The spectrum of each fluorescence signal, in fact, is partially superimposed on the scattered excitation radiation, which is much more intense. This problem is known, and is generally solved by replacing the intensity values corresponding to the superposition spectral region with zeros (FIG. 3A) or missing values (FIG. 3B).

In these figures the references F280, F340 and F429 represent the excitation fluorescence spectra at 280, 340 and 429 nm respectively, in which the intensities of the overlapping spectral regions with the Rayleigh scattering have been replaced by zeros or values. missing. These methods have drawbacks, because the replacement with zeros generates artifacts and therefore distorts the analysis by creating artificial variances. The replacement by missing values creates difficulties in the convergence of the multichannel models used for data analysis. In addition, the presence of missing values prevents the use of standardization pretreatments (see below), which can only be applied to real values. These problems are acceptable in the case of conventional techniques, based on the use of MEE of large dimensions, but become easily unacceptable in a process such as that of the invention, based on the analysis of very weakly resolved data in length. excitation wave. For this reason, the invention proposes to suppress the contribution of first order Rayleigh scattering by an innovative technique which is based on the prediction of the scattering region which overlaps the fluorescence via a generalized linear model (GLZ) with a function log binding [Davidson 1998; Rizkallah 2007]. This model is calibrated on a region of the spectrum in which the contribution of the fluorescence is negligible, and the spectral intensity is attributed exclusively to the diffusion (reference RD in FIGS. 2A and 2B). The generalized linear model is of the form: f (py) = b0 +! Ten In the equation, f (py) is the binding function of py, the expected value of y, y being the vector of Rayleigh scattering intensities which do not overlap with fluorescence, while x is a vector of indices (1,2,3, etc.) of the same size as y. Thanks to the generalized model, nonlinear relations (between x and y) can be modeled via the link function. The generalized model can be used to model dependent variables with distributions belonging to the exponential family (Normal, Gamma, Poisson, etc.). Multiple linear regression is a special case of the GLZ model which corresponds to a binding function equal to the identity function and a dependent variable (y) having a normal distribution. The parameters b; of the GLZ model are estimated by the statistical method of maximization of the likelihood (L): L = F (Y, model) = fn 1p [yi, bi]. p [y, b;] being the probability of y; conditioned by b.

The goal is to find the parameters that give the greatest probability (joint density) of realization of y for all observations. An iterative estimate (Fisher algorithm, which is a quasi-Newtonian method) is used to find the parameters b; maximizing L: alog (L) _ 0 ab, Once the b; estimated, the GLZ model is applied to the diffusion indices corresponding to the spectral region superimposed with fluorescence (fig2B - RP prediction region) to predict the intensities of "pure" diffusion. After this prediction, the complete scattering spectra (real and predicted part) are subtracted from the MEEs to obtain the pure fluorescence spectrum SF (FIG. 2C) compared to FIG. 2A showing the fluorescence spectrum partially superimposed on the spectrum of the scattered light. ). The successive operations no longer relate to the three fluorescence spectra considered individually, but to the concatenated spectra, that is to say arranged one after the other in a single column (see FIG. 4). The concatenated spectra can be submitted according to the case to: • A simple normalization by division of each emission value of each concatenated spectrum, by the norm of this same vector X.

This normalization is applied to each sample (i = 1 to I samples), ie: Xii X1J (normalized) IIXi.I where j is the number of wavelengths of the concatenated emission spectrum (j = 1 to 4545) and Xi] = • A correction by "Multiplicative scatter correction" which consists in carrying out a regression between the vector x; (the concatenated emission spectrum of each sample i) and the average concatenated vector of all the samples (i = 1 to 1 samples), then subtract the intercept and divide by the respective slope each intensity of the concatenated emission spectrum for each sample, as described below: xi = a. + bixi XI). ù (MSC) = `'See: [Bro 1998; Dhanoa 1994]

After having acquired and pre-processed the fluorescence spectra of a certain number of calibration samples, it is possible to determine the multi-channel statistical model that will be used for the analysis of other samples to be characterized. This involves calculating the vectors of

"Loadings" of this model.

We first consider the case of a multilinear model of the type

"PARAFAC" [Bro 1998]. The principle, illustrated in FIG. 5, is to decompose a three-way structure (data cube) X into a sum of external products of three vectors ("triads") a ;, b ;, c ;, plus a residue E , also in the form of a "data cube". The three lanes constituting the

"Cube" X are: samples, excitation radiation, emission wavelengths. We can write: F a bjfckf + eijk f = 1 2 xii; or Where: "i" is the index of the samples, "j" is that of the excitation radiations, "k" is the emission wavelengths, "f" is that of the F PARAFAC decomposition factors. The concatenation of the spectra makes it possible to write the PARAFAC decomposition in matrix form: Xi * JK = A (C 0 B) T; where: - 1, J and K are the numbers of samples, excitation wavelengths and emission wavelengths respectively; X, JK is the concatenated fluorescence matrix of all samples; - Be- and CKF are the vectors (at JF and KF elements, respectively) excitation and emission loadings (bilinear fluorescence profiles); A1F is the vector (at IF elements) of the "scores" (or intensities of the spectral loadings); the symbol O represents the tensor product of Khatri-Rao; and - the exponent T indicates the transpose of the column matrix. The loadings vectors are calculated from calibration samples selected to be representative of the expected total variability within the group of samples to be subsequently predicted. In other words, the model is empirical; it is therefore valid only for samples similar to those used for calibration. Parameters A, B and C of the PARAFAC model can be calculated by the least squares alternation method (non-linear iterative method). In this method a first estimate of the vector A is conditionally computed on random initial values assigned to B and C to minimize the sum of the squares of the residues. Parameter B is then updated using the estimate of A, then parameter C is updated using the new value of B; And so on. Each iterative update of A, B and C therefore improves the solution (lower error area). The algorithm converges when the improvement of the solution at the level of an iteration becomes very small (by default, this criterion is 10-). The convergence of the PARAFAC model for weakly resolved data, which is also not truly trilinear, can be problematic [Bro 1998; Harshman 1984; Rizkallah 2007]. In fact, the error surface of the model can contain local minima (points lower than their neighborhood but higher than the true minimum of the surface), col points, flat lands, and very narrow valleys that can slow down. or prevent the convergence of the algorithm.

The present inventors have noticed that imposing a limitation on the number of iterations (for example 30 maximum iterations) and / or by increasing the convergence criterion (10-2 or 10-3 instead of 10-6) improves significantly the model at the level of the parameters (loadings and scores), as will be shown below with the help of Figure 7. Indeed, an excessive number of iterations can degrade the relevance of the model. Convergence can be facilitated by imposing a non-negativity constraint since it is known a priori that the fluorescence spectra as well as their relative intensities can not take negative values. Finally, it has been observed that the pretreatment of suppression of Rayleigh scattering by application of a GLZ model facilitates the convergence of the iterative method for calculating the load vector and score vectors of the PARAFAC model. In the case of perfectly multilinear data, the PARAFAC model theoretically admits only one solution. However, in the case of a MEE limited to 3 excitations, and when fluorescence is collected at the surface, the trilinearity of the MEE matrix is strongly disturbed. As a result, the solution is no longer unique and it is necessary to develop criteria for choosing the number of factors as well as the final model.

This choice is guided by several criteria known per se [Bro 1998; Harshman1984]: - The verification of the conformity between the spectral parameters obtained (vectors B and C) and the prior knowledge of the fluorophores present in the sample analyzed. - A criterion known as CORCONDIA [Bro 1998] which serves to verify the percentage of deviation of the model compared to a perfect model multi-linearity. - The study of the percentage of variance of the fluorescence data explained by the model. - The study of the structure of the residues, which must be random. - The study of the score structure (vector A) in terms of the consistency of the evolution of the scores compared to what is expected from the application of the transformation on the batch of calibration samples. - Repeatability and reproducibility of scores, etc. The scores thus obtained by the PARAFAC decomposition of the MEEs of the calibration samples are then used to construct a multiple regression (linear or generalized) on the concentrations of the quality parameter (s) to be measured. The regression coefficients as well as the B and C vectors are stored in a model database. The model is validated by application to new known samples for which the predicted values are compared with the values measured by the reference methods. MEAs obtained from new samples are pretreated (Rayleigh removal and standardization) in the same way as the MEEs of the calibration samples. The PARAFAC model and the regression model are applied to the new samples using the stored coefficients (B, C and regression coefficients) obtained from the calibration samples. In the case of using MSC standardization, the average spectrum of the calibration samples provides the reference for correcting the spectra of the new samples. From the cube of the pre-processed matrices of the new samples to be analyzed, the new scores are calculated on the basis of the loadings vectors B and C as follows: Anouveau = (BIC) + * Xnew; where the exponent + indicates the generalized inverse of the tensor product and Xnouveau the concatenated and preprocessed spectral data of the new samples. The new scores are finally multiplied by the regression coefficients (obtained on the calibration samples) to obtain the quality parameters of the new samples. Thus the operations necessary to predict the concentrations of the quality parameters in the new samples to be analyzed consist in: - pretreat the new spectra (eliminate the diffusion and standardize); calculate the scores of the samples to be analyzed from the vectors B and C in memory; and - apply the regression equation in memory to obtain the concentrations of the parameters. The method described above was applied to predict the acrylamide content in chicory samples (n = 68). For each sample, three spectra were acquired with excitations at 280, 340 and 429 nm, pretreated by diffusion suppression and standardization, and concatenated; then the MEEs thus obtained were decomposed into four PARAFAC factors. In Figure 7: the graphs in the first column, labeled "A", represent the "scores" of the four PARAFAC factors (respectively: black dots, gray circles, squares, stars) for 100 samples; the graphs of the second column labeled "B" represent the excitation loadings of the four PARAFAC factors (curves FB1, FB2, FB3, FB4) for these same 100 samples; the graphs in the third column labeled "C" represent the emission loadings of the four PARAFAC factors (curves Fc1, Fc2, Fc3, Fc4) for these same 100 samples. The graphs in the first line refer to the case where Rayleigh scattering was suppressed by introducing missing values; those of the third line, in case it has been suppressed by introducing zeros; those of the third and fourth line, in case the suppression of the Rayleigh scattering was performed using a generalized linear model. With the exception of data that includes missing values, all other data were also standardized by MSC. The results show that the missing values in the overlap of diffusion and fluorescence (Fig.7C line 1) are poorly predicted while the zeros produce artifacts (Fig.7C line 2) consisting of abrupt variations in intensity. . Models pretreated with GLZ produce acceptable spectral parameters, with gradual intensity variations (Fig. 7C lines 3 and 4). The convergence criterion of the model of Figure 7 lines 1, 2 and 3 is 10-6 (commonly used criterion); that of the model in Figure 7, line 4 is 10 "2. The restriction imposed on the convergence of the model either by limiting the number of permissible iterations and / or by increasing the convergence criterion significantly improves the model results in terms of structure of the scores (Fig. 7A lines 3 and 4) and factor separation (Figs 7B and C lines 3 and 4) In particular, we can notice the evolution of the FB2 curve which, in the third line of the figure, shows a contribution of the three wavelengths of excitation, on the contrary, in the fourth line, the separation of the excitations is much better.Also, the shoulder of the curve Fc1 is reduced considerably when the convergence factor passes from 10-6 to 10-2 The scores from the model shown in Figure 7, line 4 were used to calibrate the chemically measured acrylamide content in the samples The chemically measured acrylamide contents s and predicted by the model based on the PARAFAC scores of the fluorescence measured on the sensor are presented in FIG. 8. In this figure, the conventionally measured acrylamide contents are plotted on the abscissa axis, those predicted by the model are on the y-axis. N is the number of samples, R is the correlation coefficient, and RMSEC is the square root of the mean squares of the errors. In particular, the high value of the correlation coefficient (almost 95%) is noted; the RMSEC is 465.47 μg / kg, ie approximately 15% of the average value of the acrylamide content measured. When the regression from the PARAFAC fluorescence scores does not lead to suitable results (for example, correlation coefficient R <0.85), another strategy that aims to improve the prediction of the regression model is applied. This method is based on multipath NPLS regression, preferably coupled to Orthogonal Signal Correction (OSC) pretreatment [Wold 1998]. The OSC pretreatment aims to reduce or eliminate the variance contained in the fluorescence signal that is orthogonal (uncorrelated) to the dependent variable (y) before performing the regression. The OSC is applied to the concatenated data after removal of GLZ diffusion and MSC standardization pretreatment. The most frequently used algorithm for OSC correction is NIPALS (nonlinear iterative partial least squares, ie nonlinear and iterative least squares algorithm). The first step is to extract the vector of the t scores of the first principal component (singular value decomposition) and orthogonalize it on the vector y: tnew = (1 - y (yTy) -1yT) t. Then, a MLR or PLS (multivariate) regression is applied to predict the orthogonalized vector: w = X + tnew (the convergence is tested on the new score t * = Xw). When t * converges, the corresponding eigenvector is calculated p = XT t * and the orthogonal effect is then removed from the fluorescence data (Xnew = X ùt * pT). In the application to the invention, a single OSC factor is calculated and then eliminated before the reorganization of the fluorescence data into a three-way structure. In the case of a single variable y, the model converges into a single iteration. Unlike the PARAFAC decomposition model, the NPLS regression model [Bro 1998] looks for explanatory factors for both X-ray fluorescence (three-way) and y (the dependent variable).

The NPLS regression therefore seeks a decomposition of the X data that results in scores that have a maximum covariance with y; an internal regression between the NPLS scores of X and y then results in the calibration: X (, x "= T (WKZWJ) T + E1; y = Tb + E2; where T is the vector of the scores of X (having the best covariance with y) and the respective loadings vectors for excitations and emissions are WK and WJ.The algorithm for the NPLS with f factors corresponding to X data (three-way) and a single dependent variable is described as follows: 1. Center the data X and y and set yo = y 2. Start the iteration f = 1 3. Calculate Z = XTy 4. Determine wJ and wK by decomposing z into singular values 5. Calculate the vectors t and calculate them in a matrix T = ~ t1 t2 t3 tf] 6. Calculate the regression coefficients b = (TTT) -1TT yo 7. Each sample X; is replaced by Xi- twJ (wK) T and y = yo - Tb 8. f = f + 1. Restart from phase 1 until sufficient explanation of y (In these equations, the index of factor fa has been omitted from t, w "wK, and b) Given that the NPLS is applied to the data after OSC pretreatment (so the greater variance perpendicular to it has been eliminated) a single NPLS factor is used for the regression. After concatenation of the MEAs from the new samples, the MSC pretreatment is applied using the average spectrum of the calibration samples as a reference. The OSC factors (p and w) calculated on the calibration samples are then removed from the new samples: X, = Xn - Xn * w * (pT * w) i * pT; 21 Xa is the matrix of MEAs concatenated and preprocessed by MSC and Xc the matrix of concatenated ECE corrected by OSC. The loadings (WJ, WK) and NPLS regression coefficients b obtained from the calibration samples are applied to the Xc matrix to predict the scores of the new samples and their dependent variable. Tä = (WJ WK) + * X ~ ÿ = Täb + e. The NPLS method was applied to the prediction of vitamin C content in green bean samples. FIGS. 9A, 9B, 10A and 10B, graphs illustrating the application of a method according to a second embodiment of the invention, based on an N-PLS model, to the prediction of the vitamin C content of various green bean samples.

After acquisition of the MEE matrices of n = 60 samples of green beans (Figure 9A: raw spectra), diffusion elimination (GLZ model) and standardization by MSC, an OSC factor was applied to eliminate fluorescence data most of the time. the orthogonal variance (non-explanatory) to the dependent variable (chemically measured vitamin C content): see Figure 9B, pretreated spectra. Then, an NPLS model with a factor was applied to calibrate the "cube" of the pre-treated MEE on the chemically measured vitamin C content. Figure 10A shows the NPLS regression coefficients for the three emission spectra (excitation at 280 nm: black line, excitation at 340 nm: gray line, excitation at 429 nm: discontinuous line). FIG. 10B shows a graph of the vitamin C contents measured conventionally (abscissa) and predicted by application of the NPLS model to three autofluorescence spectra (ordinate axis). N is the number of samples, R is the correlation coefficient, and RMSEC is the square root of the mean squares of the errors. Here again we will notice the very high value of the correlation coefficient (greater than 95%) and the fact that the RMSEC mean squared error is of the order of 15%, acceptable for the applications considered here. References: [Bro 1998] Bro, R. "Multi-way Analysis in the Food Industry Models, Algorithms, and Applications", PhD thesis, University of Amsterdam, 1998. [Dhanoa 1994] Dhanoa, M.S .; Lister, S.J .; Sanderson, R .; Barnes, R.J. "The link between multiplicative scatter correction (MSC) and standard normal variate (SNV) transformations of NIR spectra. J. Near Infrared Spectrosc. 1994, 2, 43-47. [Harshman 1984] Harshman, R. A. "How can I know if it's real? A catalog of diagnoses with three-mode factor analysis and multidimensional scaling. In: Research methods for multimode data analysis. Law, H.G .; Snyder, C.W .; Hattie, J.A .; McDonald, R.P., eds. pp. 566-571, Praeger, New York, 1984. [Rizkallah 2007] Rizkallah, J. "Chemometric Analysis of the Front Face Fluorescence Data - Estimation, of Neoformed Contaminants in Processed Foods. PhD thesis, Agro Paris Tech France, 2007. [Wold 1998] Wold, S .; Antti, H .; Lindgren, F .; Ohman. Orthogonal signal correction of near-infrared spectra. Chemometr. Intell. Lab. Syst. 1998, 44, 175-185. [Nahorniak 2003] Nahorniak Michelle L., Booksh Kart S. "Optimizing the implementation of the PARAFAC method for near-real time calibration of excitation-emission fluorescence analysis" J. Chemometrics 2003; 17; 608-617. [Rizkallah et al., 2008] Rizkallah J., Lakhal L., Birlouez-Aragon I .: "Front face fluorescence to monitor food processing and neoformed contamination" In "Optical Methods for Monitoring Fresh and Processed Food - Basics and Applications for a Chapter 4. Editor Zude M. Publisher, CRC Press, USA (2008) [Davidson 1998] Davidson, R.; MacKinnon J.G. Econometric theory and methods. "Econometric Theory and Methods," Oxford University Press 2004, pp 255-261.

Claims (12)

REVENDICATIONS1. A method for the spectroscopic analysis of at least one sample using a method for analyzing spectroscopic data based on a multi-channel statistical model, the method comprising: a) illuminating said or each sample to be analyzed (E) with a a plurality of excitation light beams at respective wavelengths; b) acquiring front fluorescence spectra (F280, F340, F429) of said or each sample, each corresponding to a respective excitation light radiation; c) pretreatment of the acquired fluorescence spectra; d) for each sample, calculating a score vector by applying to the pretreated spectra of said multi-channel statistical model, identified by a vector of fluorescence excitation loadings and by a vector of emission loadings; and e) determining the content of said or each sample in at least one substance of interest from said score vector; the method being characterized in that the average spectral difference between said excitation light beams is at least 20 nm, and preferably at least 50 nm, over a spectral range of at least 100 nm. 2. Method according to claim 1 wherein the number of excitation light beams, and corresponding fluorescence spectra for each sample, is between two and six, and preferably between 3 and 5, over a spectral range of minus 100 nm, and preferably at least 150 nm. 3. Method according to one of the preceding claims, also comprising a preliminary calibration phase comprising: i) illuminating a plurality of calibration samples by said excitation light beams; ii) acquisition of the frontal fluorescence spectra of the samples, corresponding to said excitation light rays, iii) pretreatment of the acquired fluorescence spectra; iv) determining, by an iterative method, said excitation and emission loadings vectors of the multipath model, as well as a score vector for each calibration sample; and v) determining a regression function binding said score vectors to known levels of at least one substance of interest in said calibration samples. 4. Method according to one of the preceding claims, wherein the fluorescence spectra are pretreated by subtraction of a contribution due to the Rayleigh scattering of the first order of the excitation light radiation, said contribution being calculated by means of a model linear generalized. The method of claim 4, wherein said generalized linear model is determined from a first portion (RD) of the spectrum to be pretreated, comprising only said contribution due to Rayleigh scattering, and is used to predict the diffusion contribution Rayleigh in a second portion (RP) of said spectrum in which it is superimposed on a fluorescence contribution. 6. Method according to one of the preceding claims, wherein the fluorescence spectra are pretreated by normalizing them. 7. Method according to one of the preceding claims, wherein the fluorescence spectra of several samples, calibration and / or to be analyzed, are pretreated by performing a multiplicative dispersion correction. 8. Method according to one of the preceding claims, wherein said multi-channel statistical model is a PARAFAC model, the fluorescence spectra of each sample being represented in concatenated form. 9. The method of claim 8 wherein the loadings vectors of said PARAFAC model are determined by an iterative method which is stopped when the decrease of the loss function due to the last iteration falls below a threshold value between 10-4 and 10-2.10. The method according to one of claims 1 to 7, wherein said multichannel statistical model is an NPLS model, the fluorescence spectra of each sample being represented in concatenated form. The method of claim 10, wherein the fluorescence spectra are pretreated by orthogonal signal correction. 12. Method according to one of the preceding claims wherein said excitation light beams comprise: a first radiation at a wavelength between 270 and 300 nm; a second radiation at a wavelength of between 300 and 360 nm; and a third radiation at a wavelength of between 400 and 500 nm. 14. Method according to one of the preceding claims wherein the samples to be analyzed and calibrated are samples of a product selected from: a food and a drug. 15. Use of a method according to claim 13 for measuring the evolution of the nutritional and / or toxicological properties of a food during its preparation or preservation. 15. Apparatus for spectroscopic analysis of at least one sample comprising: a set of light sources (S1, S2, S3) for illuminating said or each sample to be analyzed (E) by respective excitation light beams, presenting different wavelengths with an average spectral difference of at least 20 nm over a spectral range of at least 100 nm; means for acquiring (M, D) frontal fluorescence spectra emitted by said or each sample when it is illuminated by said excitation light beams; and a processing means (MTD) of the acquired fluorescence spectra adapted to implement a method according to one of claims 1 to 13.16. Spectroscopic analysis apparatus according to claim 15 having between two and six, and preferably between three and five, said light sources. An apparatus according to claim 16 comprising: - a first light source (S1) emitting radiation at a wavelength of between 270 and 300 nm; a second light source (S2) emitting radiation at a wavelength of between 300 and 360 nm; and a third light source (S3) emitting radiation at a wavelength of between 400 and 500 nm.