In-Line Near-Infrared Spectroscopy Gives Rapid and Precise Assessment of Product Quality and Reveals Unknown Sources of Variation-A Case Study from Commercial Cheese Production

Abstract

Quality testing in the food industry is usually performed by manual sampling and at/off-line laboratory analysis, which is labor intensive, time consuming, and may suffer from sampling bias. For many quality attributes such as fat, water and protein, in-line near-infrared spectroscopy (NIRS) is a viable alternative to grab sampling. The aim of this paper is to document some of the benefits of in-line measurements at the industrial scale, including higher precision of batch estimates and improved process understanding. Specifically, we show how the decomposition of continuous measurements in the frequency domain, using power spectral density (PSD), may give a useful view of the process and serve as a diagnostic tool. The results are based on a case regarding the large-scale production of Gouda-type cheese, where in-line NIRS was implemented to replace traditional laboratory measurements. In conclusion, the PSD of in-line NIR predictions revealed unknown sources of variation in the process that could not have been discovered using grab sampling. PSD also gave the dairy more reliable data on key quality attributes, and laid the foundation for future improvements.

Keywords: cheese production; in-line NIR; power spectral density; process analysis; sampling regime.

Figure 1

Schematic overview of the cheese production process. Control Points (CP) are marked with red circles. The in-line NIRS sensor was positioned at either CP1 or CP2, while the grab samples were taken at CP3.

Figure 2

Standard normal variate (SNV) pre-processed calibration spectra, 1/log(X) transformed for absorbance, from the in-line NIRS instrument: mean (left axis, blue) for each of the control points (drawn with different line styles) along with standard deviations (right axis, orange).

Figure 3

In-line estimates of dry matter at CP1 for a randomly selected production day in 2019. The colours indicate grouping into batches, identified based on the well-known gradient in dry matter within each batch caused by the casomatic.

Figure 4

Top: power spectral density of initial decomposed signal at control point CP1; bottom: same for signal at CP2—both from 2019. The unit is in the logarithmic scale of decibel (dB) which implies a factor 10 for each increase by 10 dB.

Figure 5

Power spectral density at CP1 for the 2022 data. The unit is in the logarithmic scale of decibel (dB) which implies a factor 10 for each increase by 10 dB.

Figure 6

Top: power spectral density of the refined decomposed signal at control point CP1; bottom: same for signal at CP2—both from 2019. The unit is in the logarithmic scale of decibel (dB) which implies a factor 10 for each increase by 10 dB.

Figure 7

Power spectral density of the refined decomposed signal at control point CP1 for the 2022 data. The unit is in the logarithmic scale of decibel (dB) which implies a factor 10 for each increase by 10 dB.

Figure 8

Zoom on a short period of the decomposition of dry matter measurements at CP1 and CP2—both from 2019. These decompositions are displayed as successive accumulations: $x_{\mu }+x_b,x_{\mu }+x_b+x_c$ etc. The CP1 period does not include the press signal, and neither include the “column” ($x_{\delta }\left[n\right]$) signal.

Figure 9

Estimates of batch means for four selected production days in 2019. Grab samples are taken from approximately every fourth batch. The estimates show the same trends, but variation is larger in grab samples.