On-line prediction of chemical composition of semi-frozen ground beef by non-invasive NIR spectroscopy

On-line prediction of chemical composition of semi-frozen ground beef by non-invasive NIR spectroscopy

Meat Science 63 (2003) 515–523 www.elsevier.com/locate/meatsci On-line prediction of chemical composition of semi-frozen ground beef by non-invasive ...

236KB Sizes 134 Downloads 87 Views

Meat Science 63 (2003) 515–523 www.elsevier.com/locate/meatsci

On-line prediction of chemical composition of semi-frozen ground beef by non-invasive NIR spectroscopy G. Tøgersena,*, J.F. Arnesenb, B.N. Nilsena, K.I. Hildruma a

MATFORSK, Norwegian Food Research Institute, Osloveien 1, 1430 Aas, Norway b Stabburet AS, 1601 Fredrikstad, Norway Received 12 March 2002; received in revised form and accepted 15 May 2002

Abstract The chemical composition of industrial scale batches of frozen beef was measured on-line during grinding by near infrared (NIR) reflectance spectroscopy. The MM55E filter based non-contact NIR instrument was mounted at the outlet of a meat grinder, and the fat, moisture and protein contents determined from the average of each filter reading throughout the grinding of the batch. The filters were selected from full spectra measurements to be as insensitive to water crystallization as possible. For on-line calibration and prediction, 55 beef batches of 400–800 kg in the range of 7.66–22.91% fat, 59.36–71.48% moisture, and 17.04–20.76% protein, were ground through 4 or 13 mm hole plates. The regression results, presented as root mean square error of cross validation (RMSECV) were between 0.48 and 1.11% for fat, 0.43 and 0.97% for moisture and 0.41 and 0.47% for protein. # 2002 Elsevier Science Ltd. All rights reserved. Keywords: Frozen meat; On-line; NIR; Near infrared; Fat; Moisture; Protein

1. Introduction Ground meats are the major ingredients in a large variety of high volume meat products like hamburgers, patties and sausages. The animal species, anatomic origin and chemical composition of the muscles, most often the fat content or lean/fat ratio, determine the basis for pricing of the ground meat cuts in the meat processing market. Ground meats are also marketed directly to the consumers with a variety of levels of fat. Thus, there is a need to analyze the chemical composition in ground meat to ensure that the consumers get the right quality. The meat processors that transform meat raw materials into consumer ready products need to be assured that they buy the right quality of meats from their suppliers. To be able to manufacture products within consumer preferences and the legislative restrictions present * Corresponding author. Tel.: +47-90-111598; fax: +47-69363851. E-mail address: [email protected] (G. Tøgersen).

in different markets, the meat composition needs to be according to specifications. Batch based or continuous processing lines with high throughput increase the demand for strict quality control and optimization of the raw material quality. A critical prerequisite to the implementation of such technologies is real-time data from the analysis of chemical composition. Before any analytical method is applied, meat raw materials intended for production of comminuted meat products are sorted by visual fat estimation after deboning. According to Kroeze, Wijngaards, Padding, Linschoten, and Theelen-Uijtewaal (2000), the ability to visually estimate fat content in meat can, to some extent, be trained, which can be used to reduce the variability of the raw materials. Analysis of chemical composition of meats has traditionally been performed off-line, by traditional ‘‘wet chemistry methods’’, some of which are accepted reference methods. Alternatives have been rapid analytical methods based on near infrared spectroscopy (NIR) or nuclear magnetic resonance (NMR).

0309-1740/03/$ - see front matter # 2002 Elsevier Science Ltd. All rights reserved. PII: S0309-1740(02)00113-4

516

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

On-line analysis has also been available for three decades using the X-ray based Anyl-Ray (Gordon, 1973). Off- or at-line applications suffer from the errors introduced by the sampling procedure as well as loss of efficiency due to time needed to carry out these procedures. Since 1996, NIR-based equipment have been available for on-line applications to determine fat, water and protein in ground beef, and suitable for commercial real life large-scale applications in meat processing. Results on the performance of such applications have been reported by Schwarze (1996) and Tøgersen, Isaksson, Nilsen, Bakker, and Hildrum (1999). Due to mismatches between supply and demand of raw meat, a substantial proportion is sold in a frozen state. Compared to fresh meats, frozen meats have the benefits of extended handling and storage time. These raw materials, especially when used in the frozen state in the first steps of a processing line, causes new challenges to on-line NIR based applications. A particular problem is measurement of meat in the transition between frozen and thawed products, which according to Forne´s and Chaussidon (1978) severely influences the interaction between NIR signals and water. Applications for on-line analysis of frozen raw meat will allow improved control of the composition of meat products. This work focuses on the development and implementation of a non-contact on-line NIR application for proximate analysis of semi-frozen raw meat, when ground through plates with holes of 4 or 13 mm in diameter. The work was mainly carried out in the processing plant intended for implementation.

2. Materials Raw material preparation for wavelength selection for the on-line instrument (MM55E, NDC Infrared Engineering, Maldon, Essex, UK) and on-line calibration on ground meat samples was done in two stages. For selection of wavelengths for the MM55 instrument filter wheel, ground beef with approximately 21 and 14% fat were mixed into a series of 11 ground meat samples, distributed in the range between these two fat levels, in increments of 0.7% fat. The mixes were blended manually and ground through plates with 4 mm holes to improve homogeneity. Ground meat, from each fat level, were filled into an open sample cup, and stored under a lid to prevent water evaporation from the surface. For on-line calibration and validation, 55 large scale (400–800 kg) batches of coarse (40–45 mm) ground beef were used. These batches were selected from the production line in a meat processing plant. The meat raw materials for each batch were prepared from 20 kg blocks of frozen meat, which had been tempered overnight to approximately 7  C prior to grinding.

3. Methods 3.1. Grinding of large scale meat batches The meat blocks intended for on-line NIR analysis were ground through an industrial scale meat grinder (Wolfking, Type C.400 Universal S 2 T. HD, Slagelse, Denmark). As a consequence of the intended use of the meat batches, 38 samples were ground through 13 mm, while 17 samples were ground through 4 mm hole plates, giving a total of 55 samples. 3.2. Sampling procedures and reference analysis Each batch was sampled five times in 1 kg samples from each corner and the middle of the rectangular batch by inserting a cylindrical (150 mm in diameter) steel tube into the meat from the surface of each point mentioned above, directed towards the bottom middle of the batch. Each of the five samples from each batch was homogenized and analyzed off-line, in duplicate for fat (Fosslet, Foss Electric, Hillerød, Denmark), protein (Kjeltec Auto 1030, Tecator AB, Ho¨gana¨s, Sweden) and water (drying at 105  C, 18 h). 3.3. NIR analysis 3.3.1. Filter selection An Infraalyzer 500 (BRAN+LUEBBE GmbH, 22803 Norderstedt, Germany) was used for off-line NIR analysis in the range 1100–2500 nm, with 2 nm increments. Each sample was analyzed at target temperatures of 5, 2, 0, +2 and +10  C. Temperatures in the samples were controlled by storing the samples in a combined freezer–refrigerator (Gastro-Line Twin, Gram A/S, DK-6500 Vojens, Denmark), in which the temperature was controllable to within  0.5  C of the target temperature. 3.3.2. On-line analysis For the on-line analysis, a MM55E (NDC Infrared Engineering LTD, Maldon, Essex, UK), equipped with a five-filter rotating filter wheel, was mounted at the outlet of an industrial scale meat grinder, as illustrated previously (Isaksson, Nilsen, Tøgersen, Hammond, & Hildrum, 1996; Tøgersen et al., 1999). The MM55E was mounted approximately 25 cm above the surface of the meat stream from the grinder outlet. The NIR signals were recorded by a laptop computer to a data file at a rate of five readings per second per wavelength, using the software Col-fast (NDC Infrared Engineering LTD, Maldon, Essex, UK). The MM55E filter wheel rotates at 20 Hz, thus each reading is an average of four revolutions of the filter wheel. The initial 25 and final 25 readings from each batch were

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

excluded, and the average of remaining readings of each filter and sample were used for subsequent calculations. 3.4. Statistics Analysis of variance was performed using Minitab1 v.13.3 (Minitab Inc. State College, PA, USA). The principal component analysis (PCA) of the full spectra measurements and PLS2 regression results for on-line data were calculated using the software Unscrambler1, ver.7.5, (Camo AS, Oslo, Norway). The calibration models for each constituent were validated by full cross validation. The prediction results were presented as root mean square error of cross validation (RMSECV): vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi u I   u X ^ 2 RMSECV ¼ tI 1  yy i¼1

where the sample number is represented by i [1,2,3,. . .,I], while y and yˆ represents the reference method value and the NIR predicted value, respectively.

4. Results and discussion 4.1. Temperature effects on the NIR spectra Acquisition of a spectrum between 1100 and 2500 nm took about 1 min on the Infraalyzer 500. Due to the temperature differences between the meat samples and the surrounding air (approximately 20  C), the surface temperature of the meat samples increased during acquisition. Therefore, surface temperatures at the time of analysis were an approximation of the target values. Fig. 1 shows the average spectrum of all 11 samples at each of the temperatures indicated. Most areas of the spectra are influenced, but the most obvious effects are the sideways shifts of the broad O–H sensitive regions around 1400–1500 and 1900–2050 nm. The effects of temperature on certain parts of the NIR spectrum of pure water have been investigated by Iwamoto, Uozumi, and Nishinari (1986) as well as Forne´s and Chaussidon (1978). Forne´s and Chaussidon (1978) analyzed transmittance spectra (1786–2326 nm) of water between 50  C and +50  C. They found the absorption band for liquid water different from that of ice. They also found distinct isobestic points for liquid water at 1957 nm and ice at 2008 nm. They explained the observed increase in absorption at lower wavelengths as effects of increased temperature, causing changes in the fractions of water molecules involved in none, one or two hydrogen bonds. Iwamoto et al. (1986) found that increasing the temperature of water from 30 to 60  C, caused a shift in the

517

absorption band around 1400–1500 nm towards lower wavelengths. They found an apparent isobestic point at 1442 nm. The shifts in the spectra were investigated further by looking at second derivative spectra around this region. The difference spectra showed absorption bands around 1418, 1466 and 1511 nm. The increase in temperature caused an increase in the absorption band around 1418 nm, at the expense of the absorption bands at 1466 and 1510 nm. Also in this case, this effect was explained as a consequence of the breaking of hydrogen bonds between water molecules resulting from the temperature increase causing changes in the fractions of water molecules having none, one or two hydrogen bonds. Sideways shifts in absorption bands also appear in the present data (Fig. 1). Isobestic points are however not that apparent. Biological tissues like meat are extremely complex matrices, and the water in meat serves as a solvent by forming hydrogen bonds with a wide variety of compounds. This causes additional molecular interactions where water is involved, and thereby more complex effects on the NIR spectra. The O–H sensitive regions also have high absorbance of NIR signals because of the relatively high content of moisture in the meat samples. In the filter based on-line application, these regions were avoided so as to increase robustness because of their sensitivity to the state of the water phase. Considering, for simplicity, ground meat as distinct particles of meat, the temperature is likely to be higher on the surface than in the interior. This is a consequence of the surface absorbance of energy from the grinding process. This results in a temperature gradient between the surface and the core of each particle as the particle leaves the grinder. Due to limited penetration the NIR signals will experience only a fraction of this temperature difference. The exact temperature that the NIR signals experience in this on-line situation is therefore difficult to measure, will be variable within a batch, and need to be solved by increasing the robustness in the calibration for industrial applications. The effects of temperature and state of water on the NIR spectra are also illustrated by PCA of each of the full spectra of the 11 meat samples measured at five temperatures. Fig. 2 shows that the samples group according to target temperatures along the first PC, while the samples within a temperature mainly are distributed along the second PC. However, meat samples at target temperature 2  C spread out between the thawed samples at 0  C and the mainly frozen samples at 5  C, along the first PC. The fact that samples of target temperature 2  C covers a major part of the first PC, can be explained by the transition of water from ice to liquid. Since the temperature in the samples could not be controlled exactly to 2  C, and the fact that the surface temperature would rise during the 1 min of data

518

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

Fig. 1. Each spectrum is the average of 11 ground meat samples with fat content between 14 and 21%, measured at 5  C (—),-2  C (- -), 0  C (– –), +2  C (—) and +10  C (- - -).

acquisition of the spectrum, a larger proportion of the water in the surface of some samples may have been in the liquid state, while others still had mainly ice in the surface. The differences in the state of the water would explain the larger spread of the 2  C samples compared with the samples at the other temperatures. The correlation coefficients between the first two PCs and the fat, water and protein contents and temperature are shown in Table 1. The first PC is highly correlated with temperature, while the second PC is highly correlated with all three chemical parameters. From Fig. 2 and Table 1, it is possible to identify wavelengths that are sensitive to changes in chemical composition which at the same time are fairly insensitive to changes in temperature. Fig. 3 contains the X-loadings for the first two PCs. X-loadings show the importance of the X-variables (NIR wavelengths) for each principal component. Large regions of the spectrum are important for PC 1, and therefore also to temperature changes or changes in the state of the water. These regions are partly overlapping regions of importance for PC2, which is highly correlated with the chemical composition of the samples. 4.2. Selection of wavelengths for on-line application When selecting wavelengths for on-line applications, several properties of the individual wavelengths and

combinations thereof must be considered. In addition to sensitivity to the parameters of interest, total signal absorbance and thereby energy reflected to the detectors as well as sensitivity to temperature changes are important. The filter combinations chosen by Isaksson et al. (1996) were modified on the basis of the full spectra measurements, using the software Algview (NDC Infrared Engineering, Maldon, Essex, UK). This software allows for simulation of filter properties such as central wavelengths and bandwidths for each filter. The simulations were performed at the applications development department at NDC Infrared Engineering. In the large-scale situation, meat blocks entered the grinder at approximately 7  C, and absorbed mechanical energy when the feeder screw forced the meat through the plates and the rotating knives. This resulted in a temperature increase in the exterior part of each meat particle. The registered temperatures rose to between 2 and 1.4  C at the grinder outlet. This temperature interval covers the phase transition from water to ice in meat, which is at about 1.8  C. The filters selected for the MM55E filter wheel should therefore be as insensitive to the effects of water crystallization as possible. To avoid the spectral effects caused by the phase transitions of water (Fig. 1) the filters were selected outside the O–H sensitive regions of 1400–1600 and 1900–2050 nm. The filters selected for the on-line application had

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

519

Fig. 2. Score plot for the first and second principal component of NIR spectra between 1100 nm and 2500 nm from 11 meat samples with fat content between 14 and 21%, measured at target temperatures of 5  C (), 2  C (*), 0  C (~), +2  C (+) and +10  C (&).

Table 1 Correlation coefficients between the first three principal components (PC) and the chemical composition and temperature of 11 samples of ground beef, measured at five different temperatures

Fat Moisture Protein Temperature

PC 1

PC 2

PC 3

0.02 0.02 0.02 0.76

0.82 0.81 0.79 0.16

0.39 0.39 0.39 0.22

central wavelengths of 1630, 1728, 1810, 2100 and 2180 nm, and each filter had a bandwidth of approximately 2% of the central wavelength. 4.3. Sampling errors and chemical analysis by reference methods A critical prerequisite for success when calibrating large batches, is to what extent the samples taken for reference analysis represent the true average of the whole batch. A nested analysis of variance was used for visualization of sources of variation between batches, between samples within batches and between replicate measurements of each sample. The latter is also given as the standard error of the reference method (Sref).

Fig. 3. Line plot of X-loadings for the first (—) and second (—) principal component of NIR spectra between 1100 and 2500 nm from 11 meat samples with fat content between 14 and 21%, measured at five target temperatures.

Some important properties of the 55 large meat batches are summarized in Table 2. The mean values and ranges appeared different for the two sub sets. The weight and temperature of each batch was not recorded for each individual batch. However, as discussed above, the temperature in such situations is extremely difficult to measure due to the temperature gradients in particles as well as the variation between particles.

520

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

Table 2 Important properties of the chemical reference data of 55 samples of ground beef, and the sub sets of batches ground through 13 or 4 mm hole plates (values are given in % of wet weight) Fat

Mean Min Max No. samples

Moisture

Protein

Total

4 mm

13 mm

Total

4 mm

13 mm

Total

4 mm

13 mm

15.91 7.66 22.91 55

18.08 11.31 22.91 17

14.23 7.66 21.10 38

64.95 59.36 71.48 55

63.28 59.36 68.91 17

66.25 60.80 71.48 38

18.49 17.04 20.76 55

18.19 17.04 19.88 17

18.73 17.36 20.76 38

Table 3 Standard deviation between batches, between samples within batches, and between parallel reference analysis within samples (Sref), for fat, moisture and protein of 55 batches of 4 or 13 mm ground beef Fat

Batches Samples Sref

Moisture

Protein

Total

4 mm

13 mm

Total

4 mm

13 mm

Total

4 mm

13 mm

4.04 0.50 0.20

3.72 0.37 0.25

4.16 0.55 0.20

3.15 0.38 0.21

2.73 0.32 0.16

3.23 0.40 0.23

0.77 0.18 0.29

0.84 0.19 0.25

0.74 0.17 0.30

Table 3 shows the amount of chemical variance between batches, between samples within batches and between replicate reference analyses. The sampling error from batches, given as the standard deviation between samples within batches, for fat and water is significantly larger (P< 0.05) for 13 mm than for 4 mm ground samples. For protein the high standard error between replicate measurements, Sref, is the most likely reason that the standard deviation between samples appears equal (P=0.42) for batches of 4 and 13 mm ground meat. Apparently, the reference analysis procedure did not provide sufficient resolution to distinguish between protein content in the two sets of batches. A low variance between samples from a batch does not necessarily ensure good accuracy in estimation of the true value of that batch average. Based on the assumption that each sample represents a random sample from the batch, a higher degree of certainty about the true mean of the batch is expected for batches with low variance between samples. However, a large variance between samples within a batch may simply indicate that the batch really is inhomogeneous, and the estimated mean could still be accurate. For prediction purposes, this only causes a problem if the recorded NIR readings fail to represent the ‘‘true’’ NIR reading of the batch, i.e. a NIR sampling error. However, for calibration purposes inhomogeneity within batches causes problems in the process of sampling for reference analysis. In this work, the average standard deviation between five samples within batches was 0.50% for fat, 0.38% for water and 0.18% for protein (Table 3).

4.4. On-line calibration and validation results Fig. 4 shows a score plot of the first two PCs for the 55 large scale batches. The 13 mm (*) and 4 mm (~) batches are partly separated in the plot. The prediction results for all 55 samples (named Total) are given in Table 4. Numbers within parenthesis represent respective values after removal of three of the 13 mm batches due to simultaneous high standard deviation between samples in the reference analysis, and large prediction error. Since these batches were not identifiable as outliers in the score plot based on the NIR-data (Fig. 4), the reason for their large prediction errors are believed to be their reference values. Based on the model for all 55 (or 52) batches, the corresponding prediction results for the subsets of 4 and 13 mm ground batches are given separately. The set of 52 batches is predicted with a RMSECV for fat and moisture of 0.83 and 0.79%, respectively, with correlation coefficients of 0.98 for fat and 0.97 for moisture. The plot for predicted vs. measured fat (Fig. 5) shows that the samples are unevenly distributed over the range. Differences in prediction errors for fat between 4 and 13 mm batches are shown from the prediction plot by the respective variation around the regression line. As expected, the prediction results for the batch sub set of 4 mm are better than for 13 mm. This is in accordance with the results from the pilot plant study by Isaksson et al. (1996), and may be explained by the differences in sampling error for reference analysis between the sub sets (Table 3). Additionally, the NIR spectra

521

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

will contain a sampling error, as discussed earlier. When the MM55 instrument measures the passing stream of ground meat, only an area of about 40 mm in diameter is measured at any given time. This represents only a fraction of the total meat stream. Therefore, the on-line application depends on the collected NIR spectra representing the true NIR spectra of the whole batch. On coarse ground meat, this would require longer grinding time, i.e. larger batches, to ensure the representativity of the average NIR spectrum. In the work of Isaksson et al. (1996), samples of 20 kg were used for analysis, and the prediction results for 13 mm ground meat were at the same level as in this work. Therefore, on batches of 400 kg and greater, the NIR sampling error is probably small compared to the sampling errors for reference analysis. In applications where coarser ground meat is used, and the batch sizes are small, applying two or more NIR instruments to the meat stream would improve the NIR sampling. The sampling error for reference analysis could be reduced by increasing the sample size or by increasing the number of samples from each batch. The difficulties of sampling for reference analysis in this work (Table 3), also apply when samples are taken for at- or off-line analysis. A sampling regime like that used in this study would be far too laborious and time

Fig. 4. Score plot for the first and second principal component of NIR spectra from 55 meat batches measured on-line with a MM55E filter NIR instrument during grinding through 13 mm (*) or 4 mm (~) hole plates.

consuming to carry out on a routine basis for every batch of meat passing through a meat processing plant. This indicates the potential benefits of on-line analysis applied directly to the process line. The prediction errors for the 4 mm batch sub set are as low as 0.48 and 0.55% for fat and moisture, respectively (Table 4). The prediction results in Table 5 are based on separate regression models for the 4 and 13 mm batch sets. The prediction errors for fat from separate models for the two batch sets are slightly higher than the prediction errors from the combined model (Table 4). For moisture this is the other way around. One would expect the 4 mm batch set to give better regression models in a separate calibration set, than in a set of all batches combined. Relatively few samples and a limited chemical range in the 4 mm batch set may have caused better models for fat from the combined set of batches. However, the random nature of the errors in the sampling system is reflected in the opposite effects seen for fat and moisture. However, for the 4 mm batch sub set, the separate models for fat, moisture and protein used only two PCs, while all other models used five PCs. Reported results on off- and at-line NIR applications on meat, are typically based on the agreement between predicted values on small samples, and reference analysis of the same sample. Typical results from such applications include prediction errors between 0.3% and 0.8% for fat (Hildrum, Ellekjaer, & Isaksson, 1995). In

Fig. 5. NIR prediction plot for 52 samples of frozen meat measured on-line during grinding through 13 mm (&) or 4 mm (*) hole plates, along with the regression lines for 4 mm (—) and 13 mm (—) samples.

Table 4 Prediction results for full cross validation on-line NIR predicted fat, water and protein in 55 (52 in parenthesis) batches of the combined or sub sets of 4 or 13 mm ground beef (values for RMSECV are given in% of wet weight) Fat

RMSECV Correlation Slope

Moisture

Protein

Total

4 mm

13 mm

Total

4 mm

13 mm

Total

4 mm

13 mm

0.97 (0.83) 0.97 (0.98) 0.95 (0.96)

0.54 (0.48) 0.99 (0.99) 0.99 (1.00)

1.11 (0.95) 0.96 (0.97) 0.94 (0.95)

0.87 (0.79) 0.96 (0.97) 0.93 (0.95)

0.59 (0.55) 0.98 (0.97) 0.96 (0.95)

0.97 (0.88) 0.95 (0.97) 0.93 (0.95)

0.46 (0.43) 0.80 (0.82) 0.68 (0.70)

0.47 (0.46) 0.83 (0.79) 0.61 (0.60)

0.45 (0.41) 0.80 (0.84) 0.63 (0.77)

522

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

Table 5 PLS2 prediction results for full cross validation on-line NIR predicted fat, moisture and protein in 17 batches of 4 mm and 38 (35) batches of 13 mm ground beef (values for RMSECV are given in % of wet weight) Fat

RMSECV Correlation Slope

Moisture

Protein

4 mm

13 mm

4 mm

13 mm

4 mm

13 mm

0.55 0.99 0.96

1.11 (1.04) 0.96 (0.97) 0.94 (0.95)

0.43 0.99 0.97

0.95 (0.86) 0.96 (0.96) 0.94 (0.95)

0.43 0.85 0.75

0.44 (0.40) 0.80 (0.80) 0.70 (0.68)

industrial settings the purpose of such applications is to estimate the composition of a large batch. The sampling error from large batches is probably much higher than these values, and is therefore the key to improving the accuracy of such estimations. The prediction results presented for on-line analysis incorporate the sampling errors from the large batches. Results for both 4 and 13 mm samples are similar to, or even lower than, the results reported on samples of 20 kg by Isaksson et al. (1996). This is encouraging for future industrial implementation because of the improved possibility to manufacture products closer to, but within specification limits. As found by Isaksson et al. (1996), the prediction errors increased with the diameter of the holes in the grinder plates. There could be several reasons for this. One is that the sampling error increases with particle size (Table 3) which will reduce the accuracy in estimating the true average of a batch. A second, but less likely, reason could be that the increased particle size in the 13 mm batches resulted in a larger variation in the raw material, which made the acquisition of a representative NIR-reading of the batches more difficult. In calibration situations, the sampling procedure and reference methods are important contributors to the calibration or cross validated prediction errors. It could be argued that given that the errors from the calibration data were normally distributed around zero, and the main sources of error originate from the reference data, the same models used for prediction of new batches would mainly be influenced by factors such as the NIR sampling error and errors caused by the regression model, and therefore could give substantially lower prediction errors in use. This would be difficult to prove without repeating the sampling at a much higher level of data resolution. However, in an experiment designed for the purpose, it would be possible to estimate the magnitude of several sources of contributing error, and therefore also estimate the remaining error that would apply to predictions. Since the filters most sensitive to water were omitted, the ability to predict water in this application is probably a consequence of the high negative correlation between fat and moisture, which in this study was

0.99. The regression coefficients for fat and water indicate the same. The univariate correlation coefficients between fat and protein were 0.88 and between water and protein 0.86. For protein, the standard error of the reference was 0.29%, which was larger than the standard deviation between samples within a batch. The prediction results for protein does not allow for implementation of these calibration models, since the prediction error was more than 50% of the total variation between batches. Thus, a predicted value with a 95% confidence region twice the prediction error on each side, would have twice the range of the batches and consequently be useless in practice. The slopes of the regression lines for protein (Table 4) below 0.77, indicate that the models would lead to large and systematic prediction errors. To improve the model for protein, the errors introduced by sampling procedure and reference methods have to be reduced. In addition, an increase in the range of protein in the set of batches, would probably improve the regression model and the calibration results. This application was accepted for industrial use by Stabburet AS, a Norwegian meat processor. Dedicated operator interface software for effective batch data and calibration management was developed. Implementation of such applications enables a more effective and accurate estimation of the chemical composition of the raw materials than off-line or at-line applications do. There is however a need for development of robust NIR instruments that allow use of larger parts of the NIR spectrum, either by use of continuous spectra, or by a larger number of wavelengths. To avoid effects of sample inhomogeneity due to sample movements during acquisition of the spectrum, on-line applications need instruments that acquire the spectrum over the entire near infrared wavelength range in fractions of a second. Development of NIR instruments is progressing rapidly. NDC Infrared Engineering have launched a new generation of on-line instruments, the MG710 which comes with 6–10 filters and a much higher rotational speed of the filter wheel than the MM55. This instrument is equipped with an IR temperature sensor, which allows for incorporation of temperature in the

G. Tøgersen et al. / Meat Science 63 (2003) 515–523

X-matrix. This would be useful in applications concerned with grinding of frozen meat. Recent developments in instrumentation suited for non-contact applications like the Corona NIR diode array spectrometer (Carl Zeiss Jena GmbH, Jena, Germany), and the low cost MicroPac Compact Spectrometer (Optical Coating Laboratory, Santa Rosa, CA, US) should be tested in on-line applications. On-line analysis of the chemical composition of both fresh and frozen raw materials enable meat processors to undertake process and quality control procedures that ensure a stable chemical composition of the end products. This also allows for effective raw material optimization, and thereby economic benefits compared to off-line or at-line analysis.

Acknowledgements This work was mainly carried out in the production facilities at Stabburet AS in Fredrikstad, Norway. The operators at the grinder and laboratory staff at Stabburet AS are thanked for the cooperation during data collection and reference analysis of chemical composition of the meat samples. We thank Robert P. Hammond at NDC Infrared Engineering for help during wavelength selection. We also appreciate the financial

523

support from ‘‘Fondet for forskningsavgift pa˚ visse landbruksprodukter’’.

References Forne´s, V., & Chaussidon, J. (1978). An interpretation of the evolution with temperature of the n2+n3 combination band in water. Journal Chem. Phys, 68(10), 1978 4667–4671. Gordon, A. (1973). Anyl-Ray determines fat/lean ratio. Food Processing Industry, 42, 495. Hildrum, K. I., Ellekjaer, M. R., & Isaksson, T. (1995). Near infrared spectroscopy in meat analysis. Meat Focus International, 4, 1995 156–160. Isaksson, T., Nilsen, B. N., Tøgersen, G., Hammond, R. P., & Hildrum, K. I. (1996). On-line, proximate analysis of ground beef directly at a meat grinder outlet. Meat Science, 43(3–4), 245–253. Iwamoto, M., Uozumi, K., & Nishinari, K. (1986). Preliminary investigation of the state of water in foods by near infrared spectroscopy. In Proceedings of the International NIR/NIT Conference (pp. 3–12), 12–16 May 1986, Budapest, Hungary. Kroeze, J. H. A., Wijngaards, G., Padding, P., Linschoten, M. R. I., & Theelen-Uijtewaal, B. (2000). Training for more accurate visual fat estimation in meat. Meat Science, 54(4), 319–324. Schwarze, H., (1996). Continuous fat analysis in the meat industry. In Report no. 96–10–1. Third European Symposium on Near Infrared (NIR) spectroscopy (pp. 43–49), 29–30 October 1996, Kolding, Denmark. Tøgersen, G., Isaksson, T., Nilsen, B. N., Bakker, E. A., & Hildrum, K. I. (1999). On-line NIR analysis of fat, water and protein in industrial scale ground meat batches. Meat Science, 51(1), 97–102.