Technical background Process control in industrial processes is increasingly important as on-line analytical equipment providing fast and reliable results becomes available. Near infrared reflection/transmission (NIR/NIT) spectroscopy is the most frequently used method in many branches of industry, while mid-infrared (MIR) spectroscopy has proved very useful for process milk analysis. Milk analysis using MIR equipment is generally more accurate than the corresponding NIR/NIT method. This is because MIR contains more specific information (fundamental absorptions) and stronger signals than NIR/NIT, which detects derived information (overtones and combination bands). In addition, full spectrum instruments based on Fourier transform infrared (FT-IR) spectroscopy for dairy product analysis in the laboratory are showing promising results with regard to the number of components (e. g. specific sugars (13), casein (7, 9), urea (6)) that can be measured. The numbers stated here refer to the list of references to be found on page 2.

Hydrolysis of lactose in milk is of interest as large racial groups suffer from lactose intolerance, i. e. they are not able to cleave lactose into glucose and galactose. Therefore, low- lactose milk products produced by the action of the enzyme ß-galactosidase are of commercial interest. The process is sensitive to the initial conditions such as temperature and ß-galactosidase concentration (2,5). Therefore, the concentrations of the sugars need to be monitored during the course of the reaction, to control the process. The reaction is typically completed within a few hours, thus it requires a fast analytical method such as MIR.

The general approach when analysing spectral data with the intention of generating future predictions of milk constituents is to use one of several multivariate methods relating the data to wet chemistry results. These methods include Partial Least Squares (PLS) regression which is described elsewhere (10).

The problem is that ordinary multivariate methods require an accurate and reproducible reference method to obtain a reliable calibration. This tends to be resource-demanding- especially when the typical number of calibration samples (fifteen to hundreds) is taken into account. In addition, during a process some intermediate species might only exist for a limited period of time, i. e. they are both produced and consumed during the reaction. In such a case it might be difficult to isolate the intermediates and to measure them using the reference methods.

Such problems can be solved using a regression method of higher order, such as PARAFAC (Parallel Factor Analysis) (1). MLR (Multiple Linear Regression), PCR (Principle Component Regression) and PLS handle second order data, as the data can be arranged in a matrix. PARAFAC requires the data to be of an order higher than two, which is the case when e. g. a spectral landscape is obtained for each sample and the whole data set can be arranged in a cube. The landscape could be obtained by measuring a reacting sample with fixed time intervals during the process. The individual spectra constituting the landscape will be related and these relationships implicitly contain information on the concentrations of all infrared absorbing compounds in the sample.

PARAFAC is able to resolve these variations and to produce concentration profiles and pure spectra corresponding to the absorbing species present in the sample. The concentrations will be arbitrary, but proportional to the true concentrations. If correlating species are present, the concentration profiles will be the sums of such correlating compounds. PARAFAC is normally the most useful method for multivariate curve resolution, as it can handle more than one sample at a time, the solutions to the mathematical problem are unique (there is only one solution to each problem), and they might resemble real spectra and concentrations. Good results have been obtained on resotving absorption and emission profiles from fluorescence spectra of sugar samples using PARAFAC (1).

In the present case, PARAFAC will not work, as the actual shape of the concentration profiles will be strongly dependent on the initial conditions of the process. In such a case, it will only be possible to analyse one landscape (sample) at a time or the unfolded data set. When a landscape is unfolded (e. g. when the spectra from the individual runs are appended to each other), one of the directions in the three-dimensional structure is lost.

Alternating Least Squares (ALS), sometimes referred to as Alternating Regression (8), is a two-way method which handles one landscape or unfolded data set at a time. Ref. (15) shows how such curve resolution methods are working in practice. ALS produces pure spectra and concentration profiles in a way similar to PARAFAC, but ALS is performed on the unfolded data set. The method has been used for resolution of infrared process data with excellent results (3, 14).

References 1 Bro, R. 1997. PARAFAC. Tutorial and applications. Chemomet. Intell. Lab. Sys. 38: 149-171.

2 Forsman, E., M. Heikonen, L. Kiviniemi, M. Kreula, and P. Linko. 1979. Kinetic investigations of the hydrolysis of milk lactose with soluble Kluyveromyces lactis ß-galactosidase.

Milchwissenschaft 34: 618-621.

3 Furusjo, E., L.-G. Danielsson, E. Könberg, M. Rentsch-Jonas, and B. Skagerberg. 1998.

Evaluation Techniques for Two-Way Data from in Situ Fourier Transform Mid-Infrared Reaction Monitoring in Aqueous Solution. Anal. Chem. 70: 1726-1734.

4 Greenberg, N. A., and R. R. Mahoney. 1983. Formation of Oligosaccharides 5- Galactosidase from Streptococcus thermophilus. Food Chem. 10: 195-204.

5 Guy, E., and E. Bingham. 1978. Properties of-galactosidase of Saccharomyces lactis un milk and milk products. J. Dairy Sci. 61: 147-151.

6 Hansen, P. W. 1998. Urea determination in milk using Fourier transform infrared spectroscopy and multivariate calibration. Milchwissenschaft 53: 251-255.

7 Hewavitharana, A. K., and B. van Brakel. 1997. Fourier Transform Infrared Spectrometric Method for the Rapid Determination of Casein in Raw Milk. Analyst 122: 701-704.

8 Karjalainen, E. J. 1989. The Spectrum Reconstitution Problem. Use of Alternating Regression for Unexpected Spectral Components in Two-Dimensional Spectroscopies. Chemomet. Intell.

Lab. Sys. 7: 31-38.

9 Kjaer. L. 1997. Say cheese-and think of direct casein determination. Scand. Dairy Inform.

2/97: 28-30.

10 Martens, H., and T. Naes. 1989. Multivariate Calibration. Wiley, Chichester.

11 Prenosil, J. E., E. Stuker, and J. R. Bourne. 1987. Formation of oligosaccharides during enzymatic lactose hydrolysis and their importance in a whey hydrolysis process: Part I : State of the art. Biotechnol. Bioeng. 30: 1019: 1025.

12 Prenosil, J. E., E. Stuker, and J. R. Bourne. 1987. Formation of oligosaccharides during enzymatic lactose hydrolysis and their importance in a whey hydrolysis process: Part II.

Biotechnol. Bioeng. 30: 1026: 1031.

13 Ridder, C., and L. Kjaer. 1995. Sweet dreams come through. Applied FTIR technology in ice cream mix analysis. Scand. Dairy Inform. 9/95: 34-36.

14 Tauler, R., B. Kowalski, and S. Fleming. 1993. Multivariate Curve Resolution Applied to Spectral Data from Multiple Runs of an Industrial Process. Anal. Chem. 65: 2040-2047.

15 Tauler, R., A. Smilde, and B. Kowalski. 1995. Selectivity, Local Rank, Three-Way Data Analysis and Ambiguity in Multivariate Curve Resolution. J. Chemomet. 9: 31-58.

In the following description the below stated abbreviations will be used: ALS = Alternating Least Squares, AR: = Alternating Regression; CVS = Cross Validation Segments, FT-IR = Fourier Transform Infrared, MIR = Mid-Infrared, NIR = Near Infrared Reflection, NIT = Near Infrared Transmission, PLS = Partial Least Squares, RMSEP = Root Mean Square Error of Prediction, SEC = Standard Error of Calibration, SEP = Standard Error of Prediction, SSE = Sum of Squared Errors; R2 = Correlation, Sr = Repeatability.

The aim of the present invention is to provide a spectroscopic method and apparatus for monitoring processes in the food industry, and specifically dairy processes, e. g. the lactose hydrolysis process. Further it is the aim to provide a method for generating calibrations without using reference analysed samples. Specifically on-line monitoring is contemplated.

Summary of the invention The present invention relates to a method and apparatus for monitoring and/or controlling a process involving food and/or beverages, e. g. proceeding in a fluid or medium in a process line or a reactor. According to the invention infrared spectroscopic analysis equipment provides spectroscopic data of the fluid or medium, the spectroscopic data covering a period of time and being evaluated by use of curve resolution, such as multivariate curve resolution, such as Alternating Least Squares (ALS), and/or Alternating Regression, (AR).

According to the invention said method and apparatus is used for determining the properties or constituents in food products, such as dairy products and/or monitoring a food process, e. g. a dairy process. Further the method may be used for generating a calibration to be used for future determining the properties of or constituents in food products, such as dairy products and/or monitoring a dairy process.

The method and apparatus used according to the invention is specifically advantageous in that no reference samples has to be measured by a reference method; the method can be used for monitoring processes, such as chemical reaction processes. Further the method can be used to recognize (or identify) unknown constituents in the sample and predict their relative concentrations.

Specifically the method can be used to provide calibrations for a spectroscopic instrument, such as an FTIR instrument arranged to monitor a reaction process such as the lactose hydrolysis process in milk. The calibrations are derived on basis of the multivariate curve resolution and the calibrations can then be stored in a data memory in the instrument and be applied for future monitoring of the process.

Brief description of drawings Figure 1 shows an infrared landscape from a chemical process.

Figure 2 is a reference versus predicted plot for lactose showing 44 calibration samples.

Figure 3 shows a reference versus predicted plot for lactose showing 23 test samples predicted using a PLS model with 5 factors.

Figure 4 shows the best 3 component ALS solution with lowest SSE (out of 100 runs) for set 1.

Upper part shows the concentration profiles, lower part the pure spectra.

Figure 5 shows the best 3 component ALS solution with lowest SSE (out of 100 runs) for set 2, Upper part shows the concentration profiles, lower part the pure spectra.

Figure 6 shows the four component ALS solution with lowest SSE for set 1. Upper part shows the concentration profiles, lower part the pure spectra.

Figure 7 shows the four component ALS solution with lowest SSE for set 2. Upper part shows the concentration profiles, lower part the pure spectra.

Figure 8 shows a lactose profile and reference lactose results plotted against each other. The profile relates to calibration set 2 with 3 components Figure 9 shows a lactose profile and reference lactose results plotted against each other. The profile relates to calibration set 2 with 4 components.

Figures 10 A and B show concentration profiles for 2 test runs (or batches).

Figure 11 shows a schematic diagram of an apparatus for carrying out the method according to the invention.

Figure 12 shows a schematic diagram of the optical system and the control system.

Detailed description The following description is given as a non-limiting example to explain the invention. An embodiment of an on-line apparatus or a system for carrying out the method according to the invention may comprise the following major components: A flow system 100, an optical MID-IR spectrometry system 200 and a control system 300 as shown schematically in figures 11 and 12.

The flow system 100, which is the object of a separate patent application, now published as WO 98/20338, may comprise the following components as shown in the example in Fig. 11: Sample intake means, comprising a tube 20 and a pump and valve means, e. g. a piston pump 40 having at least one one-way valve 21,22 at the pump inlet and at least one one-way valve 23,45 at the pump outlet. The sample intake means 20 is connected to or introduced into a process line or reactor 10 from which the samples are taken through a filter 15 and through a detachable connection e. g. mini clamps 13 comprising two flange parts and a gasket. Preferably all such connections are made according to the hygienic standards for food processing plants. The process line or reactor 10 is part of a food processing plant such as a dairy, which is not shown.

Temperature controlling means 30, preferably comprising: a preheater or cooler, e. g. a coiled steel tube embedded in or wound around an electrically heated copper cylinder, providing e. g. from 1 to 5 ml, preferably 1.5 mi of heated milk or a heated copper cylinder having an inner volume of about 1.5 ml and assigned temperature sensoring means (not shown) connected to control means 300 for controlling the preheater or cooler. The heating means 30 is designed to heat e. g. 1.5 ml milk from 1°C to a temperature about 40°C-50°C in about 25 seconds.

A high pressure pump 40, (e. g. a MSC50-h-pump as used in a FOSS ELECTRIC MILKOSCAN 50 or a single stroke pump providing a whole sample volume-e. g. 1.5 ml in one single stroke) provides the high pressure (e. g. about 400-500 bar) for homogenization of the sample. Typically at least a pressure of 200 bars is needed for homogenizing. Further the pump yield 40 will ensure a high flow rate through the IR-cuvette during a flushing period, so that the cuvette is cleaned by means of the flow rate of the milk, making further cleaning unnecessary for a number of hours. During the flushing period the pressure across the cuvette may reach 100- 200 bars. To avoid degassing in the measurement period, the pressure of the measuring branch is maintained at at least the same pressure as the pressure at the location on which the sample

is extracted from the process plant. Preferably the pressure in the measuring branch exceeds the pressure in the process plant. During a measurement the pressure is maintained at a substantially constant level by the use of a back pressure valve 88 as explained later.

In the embodiment shown in Figure 11 an in-line filter 35 provides a filtered milk passing through the measurement branch comprising the cuvette 70. Optionally a valve 45 allows the milk to bypass the filter, the milk running directly towards waste 90. The high flow rate of milk along the inside of the filter 35 will provide a cleaning of the filter 35 when the valve 45 is open. To this end the valve 45 can be controlled by the control means 300. Preferably, the valve 45 also act as a safety valve which is set to open if the pressure exceeds e. g. 400 bars.

A homogenizer 50 (e. g. a S4000 homogenizer as used in FOSS ELECTRIC MILKOSCAN 4000) is preferably included. A thorough homogenization of the liquid food product is preferred in order to obtain a representative sample (a sample containing all components in the liquid food product) inside the very thin cuvette (typically having a width of 37-50 lit). A further reason for including homogenization is that the scattering of the infrared light passing through the cuvette depends on the particle size of the liquid sample. Accordingly a uniform homogenization is essential in order to have reproducible measurement conditions. The pressure drop across the homogenizer is about 200 bars.

A further preheater or cooler 60, e. g. a coiled tube preferably wound on the periphery of a temperature stabilised IR cuvette, having an electrical resistor soldered to a copper body adjusting the temparature of the milk sample to a predetermined temperature, e. g. to about 40 °C and preferably to 50 °C before entering the cuvette, and preferably comprising assigned temperature sensoring means connected to the control means 300 for controlling the temperature of the preheater or cooler. These controls and assignments are illustrated by phantom lines in Figure 11. An IR cuvette 70, comprising a milk flow path and an IR light path crossing the milk flow path. The IR cuvette is part of an IR spectrometer allowing the analysis and/or quantitative determination of specific components of the milk in the IR cuvette.

An outlet 90 for waste and means (not shown) for collecting or evacuating the waste.

Optionally the sample may be returned to the milk processing plant.

The optical system The optical system 200 for measuring the IR absorption, preferably the MID-IR absorption, can be chosen between several known IR spectrometric systems, and realised in several ways. Prefe- rably a scanning interferometer, i. e. a FT-IR instrument is used, e. g. an MID-IR-unit as used in FOSS ELECTRIC MILKOSCAN 120 and ProcesScan FT. However, the optical system may instead include a filter wheel, comprising a plurality of IR filters appropriate for the desired measurements, e. g. as used in a FOSS ELECTRIC MILKOSCAN 50 or MILKOSCAN 102-104 or MILKOSCAN 4000 and as described in GB-B-2 028 498, EP 0 012 492 and EP 0 629 290.

A simplifie diagram of a suitable optical arrangement appears from Fig. 11 and Fig. 12.

The Box 120 is an IR-source and scanning interferometer, Scanning interferometers are well known. 140 is a detector, and 160 is a computer. 305 is communication paths between the control means and the optical means, including the path for transferring spectroscopic signals to the control system.

As an alternative to the described flow system 100 extracting samples from a process container, an IR probe (not shown) may be submerged directly into the process container. It is however the inventors experience that the use of the flow system 100 provides more accurate measurements results.

The control system The control system comprises means for controlling the flow system 100 and the optical system 200 to regularly and/or repetitively perform the measurements and means for collecting the signals from the detector 140 as well as means for converting the spectrometric signal into spectrometric data which are stored in known manner. The data are processed according to the present invention by use of software means comprising an ALS procedure stored in the processing means in the computer 160. The calculations for the determination of the quantities of the components in the milk or food product are performed in the computer 160, and they are performed by methods according to the present invention as explained in further details below.

The computer 160 may be an integral part of the control system 300 or an ordinary PC (preferably containing a Pentium) communicating with the control system 300.

In the following description a detailed explanation of a preferred method according to the invention will be given as a non limiting example, which is based on the use of Alternating Least Squares ALS. Other methods according to the invention include : Curve Resolution, such as Multivariate Curve Resolution, and more specifically e. g. Alternating Regression, AR. Simplisma ; Generalised rank annihilation method (GRAM), Evolving factor analysis (EFA); Rank annihilation factor analysis (RAFA); and Trilinear decomposition (TLD) Alternating Least Squares (ALS) relies on the assumption that the Beer-Lambert law is correct, i. e. that a spectrum (the row vector x) of a given sample can be seen as a linear combination of the pure constituent spectra (contained in the matrix A), thus x=cA [1] where c is a row vector containing the concentrations of the constituents corresponding to the pure spectra in A. In the case where more than one spectrum is measured, the general expression becomes X = CA [2]

where X is a landscape containing the spectra in its rows and C is a matrix containing the concentrations corresponding to each spectrum. In this context one sample is named X, i. e. a collection of spectra from one process run.

A typical landscape obtained during 60 minutes (with sampling every 10 minutes) from one lactose hydrolysis run with seven FT-IR spectral recordings is presented in Figure 1. Most of the variation in the spectra is in the range between 1000 and 1200 cm. This is expected, as the only compounds affected by the reaction are sugars which show strong absorptions due to stretching of the sugar C-O bonds in this range. ALS calculates the pure spectra A from the input spectra in X (the landscape) and an estimate. of C (e. g. random numbers) using a rearranged form of [2]: A = C+X [3] where C+ is some pseudo-inverse of C, followed by a calculation of a better guess of C from this A : C = XA+ [4] where A+ is the pseudo-inverse of A. The steps [3] and [4] are repeated until convergence (or a maximal number of iterations) has been reached.

Various constraints can be applied to the spectra and concentration profiles in A and C in order to avoid physically meaningless solutions (1). For example, the concentrations in C cannot possibly become negative, so a non-negativity constraint would be reasonable. In addition, when looking at compounds produced and consumed during a chemical process, the concentration profiles will be expected to have only one maximum during the course of the reaction. This leads us to applying the unimodality constraint which limits solutions to smooth concentration curves with only one maximum each. It can be argued that a non-negativity constraint will be appropriate for the pure spectra in A as well. In this specific application it is not.

As the FT-IR absorbance spectra are calculated using a water background slightly negative absorbances will result. Thus, constraining A to non-negativity will restrict the algorithm too much.

Non-negativity can be applied in various ways. The most straightforward approach is to force negative values to zero (e. g. in C) after each iteration. This is very simple and does not necessarily lead to the optimal description of X (i. e. the least squares solution). The approach employed here (adopted from The N-Way Toolbox by C. A. Andersson, Internet site: http ://www. models. kvl. dk/source/nwaytoolbox/index. htm) forces only one concentration profile (or spectrum) at a time to zero, followed by a correction of the pure spectrum matrix, A (or concentration matrix, C). This modification leads to the optimal result.

In case of more process runs contained in the same data matrix X the unimodality constraint of C would not work. In such a case only the parts of C originating from the same process should be constrained. This approach was used in the present work.

After the concentration profiles and pure spectra have been obtained the same principe can be used for prediction of the constituents in an unknown sample by applying the vector form of [4] to the spectrum of the sample : c = xA+ [5] The concentration row vector c will be in arbitrary units, but linearly related to the actual concentrations.

Calculations The data analysis and calibration is performed on the computer 160, e. g. a PC, preferably using Matlab 5.2.1 software (The MathWorks Inc., MA, USA). The pseudo-inverse in equations [3] and [4] can be calculated using the built-in functions of Matlab. The calibration routines can be taken from the PLS Toolbox Version 1.5 (Eigenvector Technologies, WA, USA).

Repeatability is expressed as a mean standard deviation (sr) of multiple determinations performed under identical conditions and is calculated as: where q is the number of samples, n is the number of replicates, Xjj is the result of the i'th replicate of the j'th sample and x is the average result of the j'th sample.

Accuracy is expressed as the Root Mean Square Error of Prediction (RMSEP) and is calculated as: N 1 2 RMSEP= V N ß (X i, reference X i epredicted) [7] where N is the number of determinations (number of samples (q) times number of replicates (n) from above) and preference and xjpredicted are the reference and predicted values corresponding to the i'th determination, respectively.

When a bias (mean difference between reference results and predictions) is observed, the Standard Error of Prediction (SEP) is used. It is calculated as: SEP = Ei-reference-X. pc. ed-bias)' [8]

If two variables are related by performing a univariate linear regression (slope a, intercept b), between instrumental responses, Xi, instrumental, and reference results, xjreference the accuracy of future predictions can be estimated by the use of the Standard Error of Calibration (SEC), calculated as: N 2 \l i=l [9] Correlation is expressed as R2, which is calculated as: N 2 1 NE (X i, reference X reference) (X i, predic (ed X predicted) 2 R2 = i=' S reference S predicted [10] where N, Xi, reterence and Xi, predicted are defined above and X reference, Sreference, X reference and Spredicted are the mean and standard deviations of the reference and predicted results, respectively.

Finally, the fit of the model to X (equation [2]) is expressed as the Sum of Squared Errors (SSE): where xi, j is an element in X, ci is as row vector containing the concentrations of the i'th sample and aj is a column vector containing the absorbencies of the j'th wavelength. M is the number of wavelengths in the spectra. Note that the reference results have not been used in the calculation of the SSE.

Experimental example verifying the applicability of the method according to the invention Sample sets. The samples obtained for this work are from New Zealand and fall within two groups: Calibration samples : This set contains 124 samples. They were collected from nine process runs (five based on skim milk, four based on whole milk) carried out in May 1997 using an experimental set-up in the laboratory. Lactozym 3000 (Novo-Nordisk, Bagsværd, Denmark) was the enzyme used. Samples were taken from the reaction mixture at various time points over a three hour period, and they were immediately heated to 80 °C in a 750 W microwave oven in order to deactivate the enzyme. Duplicate samples were taken, and the following reference

analyses and spectral measurements were carried out independently. Thus, the set of 124 samples comprises two very similar sets of 62 samples.

Test samples : This set contains 23 samples obtained from two process runs carried out in the laboratory in November and December 1997 using the same experimental set-up. Samples were taken at various intervals, and this time only the sub-sample used for reference analysis had the enzyme deactivated. (the reference analysis described here is carried out only to check the applicability of the new method, it is not included in the method according to the invention) The spectral measurement was carried out on the non-deactivated sample immediately (ie. less than one minute) after sampling in order to make the FT-IR measurements as close to an on-line application as possible. The sub-samples for reference analysis were still subjected to a heat treatment.

Reference measurements were carried out using refractive index detection.. Lactose was determined on 90 of the calibration samples, as well as the 23 test samples, using a HPLC set-up.

Spectral measurements The FT-IR spectral measurements were carried out using a MilkoScan FT 120 (Foss Electric A/S, Hillerod, Denmark, the presently preferred instrument for carrying out these measurements is however a ProcesScan FT from the same company). The infrared spectrum from 925 to 5000 cm' was recorded. The calibration samples were measured in duplicate and the test samples were measured in triplicate.

In the data analysis only the ranges 964-1542,1724-1847 and 2699-2965 cm'were used, as these are the areas containing useful chemical information.

All measurements were ratioed against water and log-transformed to give absorbance spectra. A typical time-resolved landscape for the first hour of an experiment is shown in Figure 1.

No spectral pre-processing was performed prior to data analysis, as experience shows that only minor improvements, if any, are obtained when using full spectrum data of the present type (6).

RESULTS AND DISCUSSION PLS Calibration For comparative purposes, a PLS calibration against the lactose reference results was performed. (This is not a part of the invention, it was done for comparative purposes only) The cross validated results are shown in Table 1 and a reference vs. measured plot for lactose is shown in Figure 2. Figure 2 shows the reference versus the predicted plot for lactose for 44 calibration samples (two replicates for each sample) predicted using cross validation (with 6 cross validation segments) against the reference results. The model uses 5 PLS factors with R2 = 0.996, RMSEP = 0.88 and Sr = 0.23. From these results, a model using 5 PLS factors is the best. The corresponding result om the test set shows an error (RMSEP) of 2.49. This result is shown in Figure 3. Figure 3 shows the reference versus predicted plot for lactose for the 23 test samples (three replicates for each sample) predicted using a PLS model with 5 factors. R2 = 0.987, RMSEP = 2.49, SEP = 1.55, bias 1.96 and Sr = 0.23. The reference results range from 0 to 40%

dry base lactose. The reference results and lactose predictions correlate well which is the main issue in this context.

2 CVS 4 CVS 6 CVS 8 CVS 10 CVS<BR> factors"55554 R2 0.997 0.996 0.996 0.996 0.996 RMSEP 0.82 0.90 0.88 0.93 0.88 s, 0.22 0.23 0.23 0.23 0.24 Table 1 PLS results from the cross-validated calibrations for the determination of lactose in the calibration samples. 44 samples ranging from 0 to 45 % dry base lactose were used.

'CVS = Cross-Validation Segments 2#factors = optimal number of PLS factors.

ALS Results To obtain reasonable solutions, two constraints were applied during the regression: 1. Non-negativity of the concentration profiles 2. Unimodality of the concentration profiles (except for fat) in each process run 3. Non-negativity of the spectra is a further possible constraint.

The calculations using regression are started by choosing initial values of the components. The following-very simple-start guesses were used: 1. The skim and whole milk samples had a start guess for fat of 0 and 1, respectively 2. As lactose is known to be the only sugar present in the beginning of the process, the lactose concentration of the first sample of each run was set to 1, while all others were set to 0 3. The remaining components were given random numbers as start guesses (uniformly distributed numbers between 0 and 1) As random numbers were used, many different solutions might result. For this reason the ALS run was carried out 100 times with new start guesses for each number of components.

From three to six components (corresponding to the number of pure spectra in A, equation [2]) were tried on both (almost identical) calibration sets of 62 samples. When six components were used, the concentration profiles and pure spectra became noisy and highly correlated, so they will not be discussed in the following text.

The best models, in terms of how well the X matrix is described (measured as SSE) and how well the"lactose"profile correlates with the reference results (measured as R2), are shown in Table 2 (A and B parts).

3 components 4 components 5 components Set 1 Set 2 Set 1 Set 2 Set 1 Set 2 SSE 0. 56 1. 01 0. 38 0. 44 0. 26 0. 47 R2 0. 928 0.942 0.942 0.971 0.978 0.977 SEC 3.47 3.15 3.12 2. 23 1.92 2.00 A: Results of the calibration sets (62 samples each)-the models selected on the basis of the lowest SSE.

SSE 0. 71 1. 06 0. 44 51. 63 0. 40 10. 50 R2 0.928 0.942 0.971 0.986 0.991 0.992 SEC 3.47 3.15 2.22 1.54 1.26 1.16 B: Results of the calibration sets (62 samples each)-the models selected on the basis of the highest correlation (R2) to the lactose reference results.

R 2 0. 894 0. 891 0. 959 0. 980 0. 959 0. 940 SEC 3.46 3.50 2.16 1.51 2.14 2.61 s, 0.15 0. 15 0.16 0.27 0.16 0.13 C: Results of the test set (23 samples) using the models (selected by use of the SSE) obtained from the individual calibration sets.

R 0. 894 0. 891 0. 959 0. 953 0. 978 0. 973 SEC 3.46 3.50 2.15 2.29 1.59 1.78 s, 0.15 0.15 0.16 0.14 0. 19 0. 19 D: Results of the test set (23 samples) using the models (selected by use of R2) obtained from the individual calibration sets.

Table 2 ALS results of the two calibration sets (part A and B) and the independent test set (part C and D) using 3,4 and 5 components. The reference results range from 0 to 45 % dry base lactose.

The results for three to five components were as follows: With three components, the same pure spectra and concentration profiles were reached every time on both calibration sets-at least the differences were insignificant. The solutions with the lowest SSE are shown in Figures 4 and 5. Note that both pure spectra and concentration profiles have been normalised to make a presentation on the same scale possible. The pure spectra describing fat (having strong absorptions in the high end of the spectrum due to C-H stretching vibrations) have very negative contributions in the areas where the sugars absorb, reflecting that high-fat samples generally contain less lactose as a result of the displacement of the water phase by fat.

With four components many different solutions were reached. They belonged to a limited number of groups of solutions inside which the variations were small. Some results are shown in Figures 6 and 7. Of the four components at least two are sugars (having strong absorptions in the 1000 to 1200 cm-'range), one is fat (strong absorptions between 2800 and 3000 cm~'), while the last one is difficult to assign. Of the sugars, the component decreasing rapidly through each batch is lactose.

With five components the problem of finding the optimal solution becomes even more difficult. But, as is evident from Table 2, good lactose correlations were still obtained. As there is no major improvement when going from four to five components, the three-and four-component solutions were chosen for further examination.

An ALS with three components gives the most stable result and the most reasonable pure spectra, but the correlation to the actual lactose concentration is relatively poor. In Figure 8 the lactose concentration profiles are plotted against the reference results, and the resulting plot is clearly non-linear. R2 is 0.942 and SEC is 3.15. Figure 9 shows a similar plot for an ALS on calibration set 2 with 4 components. R2 is 0.971 and SEC is 2.23. The four-component model (Figure 9) gives a much more linear reiationship to the lactose reference results. It cannot be due to overfitting (i. e. a too optimistic estimate of the error) as the reference results were not involved in the optimisation. In addition, the concentration profiles of the four-component model agree with previous observations that not only the monosaccharides (galactose and glucose), but also various other sugars (containing two or more monosaccharide units), generally known as the oligosaccharides, are formed during the process (4,11,12). The shapes of the concentration profiles are actually very similar to these previous observations. The way in which these sugars are distributed among the last two components (Figures 6 and 7) can vary between different ALS runs, which is the reason for the many different solutions seen when four (and five) components are tried.

The SSE and/or the R2 may be used as the selection criterion. When three components are used, both criteria give almost the same result, while the SSE criterion gives a somewhat higher prediction error in the case of four and five components. In both cases there is a significant improvement over the three-component result. As use of SSE and R2 give similar results SSE is preferred as it does not involve reference data.

The final test of the new method was done by using the ALS models on the test set obtained from two new process runs. The results are shown in Table 2 (C and D parts). There is a large improvement in SEC when going from three to four components. It is almost independent of the method (R2 or SSE) used for selecting the optimal model. These results lead to the conclusion

that there is nothing gained by selecting the optimal model by use of the lactose reference results (i. e. by looking at R2). In fact, the best result is obtained by using the SSE. This is a very promising result, since there is no need for reference analyses when process data of the present type are analysed. ALS alone can be used for generating a model which can be used for future process monitoring. Note that the ALS concentrations are in arbitrary units, so only relative process changes can be detected.

The results in Table 2 should be compared to the PLS result shown in Figure 3. It is seen that the ALS predictions of lactose (using four components) are almost as good as the PLS predictions-in some cases the results are the same.

The normalised predictions of three of the four components (using the model with lowest SSE based on calibration set 2) for each of the two processes included in the test set are shown in Figure 10 A and 10 B together with the corresponding results from the PLS model and the reference method. Fat is omitted, as it is constant during each run. The ALS lactose predictions do not agree perfectly, neither with the reference nor the PLS results based on the same spectra, but they follow roughly the same curve. The main reason for the disagreement between the ALS and PLS results are the negative PLS lactose predictions which are due to the earlier discussed bias of the PLS calibration. Both PLS and ALS concentration profiles follow a smooth curve, which should be expected when dealing with a chemical reaction. Thus, the fluctuations in the lactose reference results are likely to be caused by the lack of reproducibility of the reference method rather than real variations in the lactose content.

The two test runs (Figure 10 A and 10 B) gave the same shapes of the concentration profiles of the third and fourth component as seen during calibration. Therefore Component 3 is assigned to the sum of galactose and glucose, and Component 4 is likely to be caused by oligosaccharides formed during the reaction. Another data set with reliable reference results on other sugars is required to confirm this.

The remaining problem allowing implementation of ALS for practical use in dairy process monitoring is how to select the optimal number of components in the ALS model. This corresponds to the problem of selecting factors in PLS, but in the ALS case there is no prediction error (e. g. RMSEP) to minimise. In the present case the obvious choice would have been three components, as this gave the most stable result. Only the comparison of the profiles to actual lactose results indicated that four components were optimal. Methods for determining the number of independently varying species present in the samples are therefore required.

Scores and/or loadings obtained through Principal Component Analysis (PCA) (10) could solve the problem. In the present case, when performing a PCA on all calibration samples,

the scores (not shown) contain structure (originating from the batch structure of the data) revealing up to four or five components. Thus, four or five components would be expected to be optimal in ALS, which supports the actual findings shown above.

In this case it was possible to extract concentrations from FT-IR data from a lactose hydrolysis process and to monitor new process runs using this knowledge. The present data set must be considered to be worst case, as the infrared spectra of reactants, intermediates and products are very similar, i. e. they are all sugars which give roughly the same absorption peaks.

Resolving spectra from processes where the compounds are much less similar should therefore be easier, especially the problem of determining the number of components should be less difficult.

Conclusions It appears from the foregoing that ALS is a promising method for use in dairy process optimisation. Without the need for reference analyses it is possible to extract e. g. four components from lactose hydrolysis process data (fat, lactose and two other sugar components) and to obtain a lactose prediction error similar to the one obtained from an ordinary PLS regression. Such use of ALS for reference-independent prediction of process parameters is not limited to dairy products only, but is likely to be useful for process monitoring and identification of intermediates in all branches of the food and beverage industry.

By use of ALS combined with FT-IR it becomes possible to obtain quick information on compounds present during the process, but not necessarily by the end of it. A further advantage (in many cases the most important) is that the pure spectra obtained by ALS makes it possible to generate predictions of process parameters without the need for expensive and time consuming reference procedures.

While a few particular embodiments of the invention have been mentioned above, it will be understood, of course, that the invention is not limited thereto since many modifications may be made, the monitoring may performed by collecting data at regular time intervals or at intervals defined in other ways, eg. by certain recognisable changes in concentrations, the process could be e. g. a batch reaction, a dairy process, a distillation or fractioning process, a fermentation process or other reaction process involving food feed or beverages and it is, therefore, contemplated that the appended claims shall cover any such modifications as fall within the true spirit and scope of the invention.