NIR DETERMINATION OF FLOUR BAKING PROPERTIES
Technical field
The present invention relates to a method for the determination of flour baking properties. More in particular the flour baking properties are modelled by analysing samples from many different flours using NIR-spectroscopy, whereupon multivariate analysis is used for determining the correlation between measured reflectance spectra and the volume of bread baked from the different flour samples, which yields a calibrated model by means of which the volume of a bread, having been baked from a flour that was not part of the calibration, can be predicted by analysing the flour using NIR-spectroscopy.
Background of the invention
In order to determine the quality properties of a flour many different parameters can be examined, for example ash contents, falling number, sedimentation and protein content. In the baking industry there is a constant need of determining the baking properties of a specific flour, whereby practical tests have shown that the above-mentioned parameters do not describe the flour baking properties in a satisfactory way. In order to determine a certain flour baking property, such as the volume of a bread baked from a certain amount of the flour, currently the flour has to be subjected to test baking according to a standardised procedure
The determination of flour baking properties using test baking is not a particularly accurate method; among other things there are a number of factors, such as humidity and temperature, which are difficult to control but have an influence on the baking result. The treatment of the dough can also, despite the standardized procedure differ between bakers, which also affects the reliability of the test baking result. However, above all, test baking is an expensive and time demanding procedure, and can therefore not be performed as often and as continuously as really needed. For example in a flour mill a method by means of which the baking properties of a specific flour could be determined on line would be an extremely
valuable instrument, which could be used for continuous adjustments of flour mixtures, and make possible a more optimal pricing, i.e. the better the baking properties of the flour is, the more the grain producers will get paid, and the more the users of the flour (e.g. bakeries) will pay.
Prior Art
In NIR-spectroscopy electromagnetic radiation at wavelengths in the near infra-red region are directed towards the target. The impinging radiation is partly absorbed, and the radiation energy will give rise to vibrations in different atomic bonds in the target. The radiation that is reflected from the target therefore contains information about the composition of the target. In analysis using NIR-spectroscopy wavelengths of the emitted radiation is commonly given discrete values over an interval, and at the same time it is measured how much of the radiation that is reflected at the different wavelengths. In this way a so called reflectance spectrum is obtained. NIR-spectroscopy is a well-known technique and a more detailed description will therefore not be given here, for further details see e.g. Skoog-Holler- Nieman, Principle of instrumental analysis, chapter 17D, 5:e edition 1998.
The relation between the composition of a flour and its baking properties is complex and depends on many different factors, which is reflected by the above-mentioned circumstance, that it is not sufficient to determine e.g. the protein content of a flour in order to know its baking capability. For NIR-spectroscopy this means that it is not possible to directly derive e.g. the final volume of a loaf baked from the flour, by measuring the reflectance from a flour sample at certain wavelengths. Instead the bread volume should be determined using a reference method, in practice test- baking, and correlate the obtained NIR-data to this bread volume. The correlation between NIR-data and bread volume is used as a calibration of the method, and the volume of a bread, baked from a flour that is not part of the calibration can therefore be predicted by measuring the reflectance of the flour at the same wavelengths as those used for the calibration.
For all practical purposes, the measure "bread volume" (predominantly used in Sweden) can be substituted by the measure "loaf height" (predominantly used in other countries, e.g. Great Britain). The two measures will yield the same predictions for an unknown flour, if a model is built in accordance with the invention based on either measure.
The idea of using NIR-spectroscopy to determine baking properties of flours, is not completely unknown, see e.g. an article by B.G. Osborne in J.Sci. Food Agric. 35, pp. 106-110, 1984. However, the experiments have not been particularly successful and today it is common opinion that the bread baking capability cannot be determined with NIR-spectroscopy with better results than what can be anticipated knowing the protein content. An article in Cereal Chem. 64, pp. 407-411. 1987, G.L. Rubenthaler and Y. Pomeranz shows that a reasonable correlation can be obtained between NIR-spectra, measured for flour samples at three different wavelengths, and the volume of a loaf baked from the flour. However, later investigations have shown that this good correlation probably was the result of specific selection of samples, which varied substantially with regard to protein content and bread baking capability, see S.R. Delwicho and G. Weaver, Journal of Food Science 59, pp. 410-415, 1994.
In the last mentioned document which is regarded as the closest prior art, measurements of the reflectance for different flour samples over the wavelength range 1100 nm to 2498 nm in steps of 2 nm by using NIR-spectroscopy are reported. Correlation between the measured spectra and the height of a loaf baked from the flour is determined, using both principal component analysis and partial least squares regression. Whether or not any differentiation of data is performed before the analysis is not derivable from the document. The conclusion from the experiment is that the possibility for NIR-spectroscopy for the determination of bread volume is small.
Summary of the invention
Thus, there is a need for a method by means of which the baking properties of a flour can be predicted or determined easier and quicker than what is possible with the prior art methods. A conceivable method is near infra-red (NIR) spectroscopy, today used e.g. in food analysis and agricultultural analysis, and for quality control in the industry. Examples of food analyses performed using NIR-spectroscopy is the determination of water content and protein content in wheat flour.
The object of the present invention is therefore to provide a method for the rapid and simple prediction or determination of properties of a flour, by means of which flour samples are analysed using NIR-spectroscopy, and wherein the correlation between the measured spectra and properties of the flour determined by means of other methods, is determined using multi-variate analysis. The result of the analysis is a calibrated model, by means of which the different properties of a flour is predicted considering only the NIR-spectrum of the flour.
The possibilities of using NIR-spectroscopy for the determination of flour baking properties, expressed e.g. as the volume of a bread baked from the flour, are substantially improved over prior art by using the inventive method, which exhibits the features defined in claim 1.
Thus, the invention relates to a method of flour analysis, for determining a baking property thereof, comprising the steps of test baking flour from a number of flour samples to obtain bread, and measuring at least one property of the bread obtained from each flour sample, measuring the reflectance for each flour sample at a plurality of wave lengths in the wave length range 400 - 2500 nm. The reflectance data for each flour sample are stored. Multivariate analysis is used to create a calibrated model of the relation between the NIR spectrum and the measured property of said bread obtained from each respective flour sample. The multivariate analysis comprises Partial Least Squares Regression, and an exclusion of wavelengths at which the reflectance values from the flour do not contain any
significant information about the property to be predicted. The baking property for a bread baked from an unknown flour, is predicted by using said calibrated model. Some of the above mentioned steps can be excluded under certain circumstances, which will be apparent from the detailed of the invention given below.
Brief Description of the Drawings
The invention will be describe in detail below with reference to the drawings, in which Fig. 1 shows the correlation between bread volume and NIR spectra;
Fig. 2 shows the correlation between bread volume and NIR spectra, using MSC and differentiation;
Fig. 3 shows the correlation between bread volume and NIR spectra as in
Fig. 4 with the additional use of Jack Knife exclusion of data;
Fig. 5 shows regression coefficients for wavelengths without Jack Knife exclusion; and Fig. 2 shows regression coefficients for wavelengths with Jack Knife exclusion.
Detailed description of the invention
In the method according to the invention use is made of a number of methods known per se for the analysis of data. These need therefore not be explained in any closer detail, but a very short description will be given below.
Principal component analysis (PCA) comprises splitting up a data set in a number of principal components, also known as eigen-vectors, each of which describes a part of the variation in the data set.
Partial least square regression (PLS or PLS-regression) is a type of multiple regression where the components are determined both by the dependent and the independent variables.
Cross-validation means that the same samples are used for building a model and testing, which is carried out by excluding one sample at a time from the calibration data set, while calibration is performed with the remaining samples. The excluded sample is predicted and prediction residual is calculated. The procedure is repeated until all samples in the data set have been excluded once, whereupon all residuals are combined for the calculation of RMSEP and the correlation coefficient, see below.
The method according to the invention begins by selecting a number of flour samples. These samples are analysed with respect to the quality parameter(s) of interest. Examples of such parameters are falling number, which is a measure of the alpha-amylase activity, ash content, which is a measure of the fraction of shell parts of the flour, bread volume, which is the volume of a bread baked with a certain amount of the flour according to a standardized procedure, and protein content. In the present description only the correlation between the various samples, NIR-spectra and the bread volume will be discussed, but of course there is nothing to prevent that any quality parameter be studied. In fact the inventive method can be used to study the correlation between the NIR-spectra of the flour samples simultaneously with one, two or more quality parameters.
Other baking properties that can be determined with the inventive method are e.g. texture, toughness, porosity, just to name a few.
The selected flour samples are analysed by using NIR-spectroscopy, whereby a reflectance of the samples at different wavelengths is measured. Hereby so called reflectance spectra are obtained, which are stored as data tables in a suitable media for further analysis. In the measurements it is of course extremely important that wavelengths, the reflectance of which is characteristic for the quality property to be studied are covered. In other words, if e.g. the content of a certain protein has a
large impact on the bread volume, it is important that the reflectance spectrum comprises wavelengths indicating the presence (or absence) of this protein. An important feature of the method according to the present invention is that in the analysis such wavelengths are excluded which do not contain any information with respect to the quality property looked for. Thus, this means that a large wavelength range can be covered by close measuring steps without having a negative influence on the analysis results due to a large amount of "noise". Surprisingly it has been discovered that with said features of the inventive method an unexpected good correlation between the quality parameters, e.g. bread volume, and NIR-spectrum is obtained, compared to prior art methods.
Before analysing the measured reflectance spectra, suitably so called multiplicative scatter correction (MSC) of the stored reflectance data tables is performed. MSC is a well-known transformation method, used to compensate for any (unwanted) light scattering that may be at hand, giving rise to additive and/or multiplicative effects, which could dominate the information in the data tables. MSC minimizes differences in light scattering between flour samples, that can depend on different packing densities in the samples or differences in the particle size of the flour samples.
In order to reduce the influence of variations in the base line between sample spectra, suitably a differentiation of the data is performed. The method used was Norris- differentiation for the calculation of the first order derivatives, but other methods are also applicable, for example differentiation according to the Savitzsky- Golay-method.
The need for using MSC and/ or differentiation is determined by the experimental circumstances prevailing in the measurements. If the light scattering is neglectable and furthermore the same between samples, MSC need not be performed. If the variation in the base line is neglectable between spectra, no differentiation needs to be done.
Before any final analysis of the reflectance data of the flour samples is performed, it can be appropriate to perform a principal component analysis (PCA), whereby the stored data- tables (spectra) are divided into a number of principal components, each of which describes a part of the variation in the spectra. In order to identify so called outliers, i.e. samples which for some reason or another deviates from the others, it is suitable to project data for each sample in a new coordinate system, using for example the two first principal components that explains the major part of the variation of the data tables as coordinates. I should be noted that in this analysis it is thus only the variation in the reflectance spectrum of the flour samples that is analysed, not how well this variation correlates with variations in the quality properties of the samples. The main purpose of the principal component analysis is to get a picture of those variables that contribute the most to the variations in the data tables, and in which way they contribute and also to quantify the amount of usable information in the data tables.
The major analysis of the data tables is then performed using partial least square regression (PLS). In PLS it is assumed that the dependent variable can be described by a linear combination of a number of independent variables, and a regression equation is proposed:
Y= a+bιXi+b2X2+...+bnXn Equation 1
where Y is the independent variable, in this case bread volume, a is a constant, Xi is the principle component I, where 1 < i > n, and bi the regression coefficient for the principal component Xi. In practice the analysis is performed using a suitable program package, which determines both number, n, principal components, and which these principal components are. Normally n is between 5 and 20. It should be noted that it is not necessarily the same principal components that are obtained in PCA as in PLS since the latter also considers the independent variable (bread volume). The program package also determines the constant a and the regression coefficients bi, b2, ..., bn. The result of the PLS analysis is therefore a calibrated model, by means of which the bread volume Y can be predicted.
As a measure of how well the model describes data it is common to give the correlation coefficient, r, and RMSEP (Root Mean Square Error of Prediction), which can be calculated in accordance with:
Equation 2
and
where yϊ are the measured values of bread volume for the n different flour samples and Yi are those predicted by equation 1.
In this connection it should be mentioned that the regression equation above represents the underlying model, which normally is not explicitly presented to the user if a commercial program package is employed. That is, the user will never see values of the constant a, coefficients bi or the principal components Xi. These will be stored as a calibration, which then can be used to predict bread volume from a measured NIR-spectrum.
As mentioned above, wavelengths the reflectance of which do not contain any information about the property looked for should be excluded in the analysis. This procedure improves the reliability of the PLS model substantially, at the same time as it enables scanning a larger wavelength range without the PLS regression being negatively influenced by reflectance data not containing any information about the properties sought for.
Below the results of an experiment is presented, in which the above described procedure is used to determine baking properties of flour. In the experiment only 40
wheat flour samples were included, but it is to be understood that the method according to the invention advantageously can be used for a much larger number of samples. If as an example, the inventive method is used at a bakery or a flour mill, suitably reflectance spectra for flour samples, the respective bread volumes of which are determined by test baking, can be measured daily during a number of years in sequence. In this way data for different flour samples, having different baking properties, are included in the calibration of the model, which improves the possibility to predict the bread volume for a new flour, in the experiment below measurements are presented in the wavelength region 400-2498 nm, but of course the inventive method can be used for other wavelength ranges.
Preferred embodiment
The above described method was tested by analysing 40 heavily milled wheat flour samples. The samples, including values of their respective bread volume, were provided by Nord Mills AB and originated both from spring and autumn harvest 1999.
For the measurements of the reflectance of the samples an NIR-instrument, NIRS- System Model 6500 from Foss-Tecator was used. The wavelength range was 400- 2498 nm and the reflectance was measured in steps om 2 nm, i.e. for each sample 1050 measurement values were obtained, which were stored as data tables for further analysis. The analysis was carried out using so called double samples, i.e. each sample was subdivided in two parts and the analysis was performed separately on the two samples.
For the analysis the program The Unscrambler® version 7,5 from Camo AS was used. Before the measured reflectance data tables were analysed, data were transformed using MSC and Norris-differentiation for the purpose of compensating for unwanted light scattering and variation in base line respectively, as described above. The analysis started with a PCA, whereupon the samples were plotted in the diagram with the two first principal components as coordinate axes, whereby it was noted that one sample differed much from the others. This sample, having an
extremely low value of bread volume, was therefore discarded from the further analysis. The correlation between NIR-data and bread volume was calculated using PLS-regression, whereby wavelength containing no information regarding bread volume were excluded/deselected. This exclusion is performed automatically if a function in the program the Unscrambler® called "Jack Knife" is selected.
The use of "Jack Knife" for the elimination of useless variables is discussed in an article by Martens et al in Food Quality and Preference. 11 (2000) pp 8-9, 2.8, which is incorporated herein by reference in its entirety.
The result, shown in figure 1, was a good correlation, r2=0.789, between bread volume and NIR spectra. This can be compared with the result from corresponding experiments according to prior art, where the correlation was practically nonexistent, r2=0.209.
The error for the method was about 90 ml, RMSEP = 94.21492, which is to be compared with the bread volumes that were found in the range 2000-2500 ml, and the reproduceability of the test baking i.e. approx. 50-70 ml.
Figs. 2 and 3 illustrate the improved correlation when Jack Knife is used.
Figs. 4 and 5 illustrate spectra before (Fig. 4) and after (Fig. 5) Jack Knife exclusion.
Even if the invention has been described above with reference to a particular, preferred embodiment, the result of which is presented in the appended figure, it is obvious for the skilled man that variations and modifications are possible within the scope of the inventive concept such that it is presented by the description and the appended claims.