Application Note 79: Analysis of Moisture in Skim Milk Powder using the NIT-38 Dairy Analyser

April 2006

Introduction:

A calibration was developed for measuring Moisture, Fat and Protein in Low fat Milk Powder. The original calibration set contained approximately 20 samples of milk powders. The reference analyses were based on the Dairy America laboratory in California.

To test the calibration, 2 sets of 10 samples were analysed using the original calibration.

Table 1. Moisture Prediction Data using Original Calibration

Sample ID	NIT-38	Ref		Diff
201s1	3.37	3.99		0.624
202s1	3.38	4		0.623
203s1	2.94	3.74		0.8
204s1	3.09	3.92		0.832
205s1	3.47	4.13		0.662
206s1	3.23	3.86		0.632
207s1	3.05	3.71		0.66
208s1	3.32	3.99		0.669
209s1	2.98	3.73		0.748
210s1	3.13	3.78		0.652
211s1	2.95	3.78		0.831
212s1	2.80	3.7		0.896
			SEP	0.10
			Bias	0.72
3201s1	2.57	2.97		0.399
3202s1	3.48	3.48		0.004
3203s1	2.99	3.16		0.171
3204s1	3.86	3.94		0.085
3205s1	3.62	3.68		0.064
3206s1	3.42	3.46		0.036
3207s1	3.72	3.72		-0.002
3208s1	3.38	3.51		0.129
3209s1	3.41	3.61		0.201
3210s1	3.45	3.62		0.168
			SEP	0.21
			Bias	0.18

Figure 1.

Table 1 and Figure 1. show the results of the test. It can be seen in figure 1. that there are two lines representing the first set 201-210 and the second set 3201-3210. It was noted that the first set was based on UDA Oven Drying method and the second set is based on Dairy America’s laboratory.

The difference between the two sets of data can be explained as either due to inter laboratory bias or an effect causing the NIT-38 to shift results on different days. Sample or instrument temperature differences are possible causes of the shift in results. The records stored in the NIT-38 show that the instrument temperature was constant, ie, 30 and 31 C for the two sets. Unfortunately we do not have a record of the sample temperature. It should also be noted that there has not yet been any temperature stabilization samples added to the calibration set.

Without the data on sample temperature, I chose to correct the interlab bias by adding 0.59 to the first set of Ref data. The 0.59 was the average difference in bias between the NIT-38 Moisture and the Ref values for each set. This was done to try and bring the two sets of laboratory data in line with each other.

A feature of NTAS (NIR Technology Australia Software) is the ability to test a calibration using an independent sample set. As such, reloaded the original calibration set into NTAS and used the 201-210 and 3201-3210 samples(after removal of the interlab bias) as a prediction set. Table 2. and Figure 2. shows the results of predicting these sets and selecting the optimum number of PC in the calibration.

In the original calibration, 9 PC’s were used, however the optimum number of PC’s proved to be 6 based on the prediction set. Note that the SEP(Standard Error of Prediction) was 0.1% and the bias 0.03%.

Table 2. Moisture Prediction Data

Sample ID	NIT-38	Ref		Diff
201s1	3.36	3.4		0.04
202s1	3.20	3.41		0.21
203s1	3.20	3.15		-0.05
204s1	3.21	3.33		0.12
205s1	3.30	3.54		0.24
206s1	3.29	3.27		-0.02
207s1	3.26	3.12		-0.14
208s1	3.25	3.4		0.15
209s1	3.18	3.14		-0.04
210s1	3.27	3.19		-0.08
211s1	3.03	3.19		0.16
212s1	3.13	3.11		-0.02
3201s1	2.93	2.97		0.04
3202s1	3.50	3.48		-0.02
3203s1	3.15	3.16		0.01
3204s1	3.87	3.94		0.07
3205s1	3.79	3.68		-0.11
3206s1	3.42	3.46		0.04
3207s1	3.68	3.72		0.04
3208s1	3.52	3.51		-0.01
3209s1	3.60	3.61		0.01
			SEP	0.10
			Bias	0.03

Figure 2. Plot of NIT-38 Moisture vs Ref Moisture

New Calibration

The next step is to add the new spectral data to the original calibration set and to develop a new calibration. Figure 3. shows the plot of the NIT-38 Moisture vs Ref Moisture for the combined spectral set.

Figure 3. Calibration data

To see if the new calibration provides an improvement, the model was used to predict the 201-210 and 3201-3210 samples. Figure 4. shows the plot of the prediction data using the new calibration. The SEP is reduced to 0.08%.

Figure 4. Moisture Prediction Data

Conclusion:

The new calibration appears to be more robust and more accurate. This is often the case because the selection of the number of PC’s used in the calibration is now based on prediction data rather than just calibration data.

Further sets of samples should be analysed using this new calibration. As well, temperature stabilization samples should be scanned and added to the calibration set.

The development of robust and accurate NIR calibrations is a gradual process. The objective is to develop a cause and effect relation between the NIR spectral data and the reference data. Initial calibrations can only be considered as starting calibrations so that further data can be collected which represents the true variation in samples, environment and instrument conditions. This is what we can see from the above data.