TY  - THES
T1  - Self-learning modules for spectra evaluation
A1  - Sadeghi Vasafi,Pegah
Y1  - 2022/04/05
N2  - Monitoring milk processing is an essential task as it affects the quality and safety of the final product. The aim of this investigation was to develop and analyse the self-learning system for the supervision of the processing of milk. In the self-learning evaluation module, several algorithms for data analysis of near infrared (NIR) and Raman spectra was implemented for the prediction of sample quality and safety.
In the first part of this thesis, the use of NIR spectroscopy for controlling milk processing was investigated. For this reason, a high-quality quartz flow cell with a 1 mm pathlength including temperature controlling option for liquids was implemented. For sample preparation, UHT-milk with 1.5 % fat content was measured at 5 °C and considered as the reference milk. Samples with various changes such as added water and cleaning solution, different fat content and 
temperature as well as milks from various suppliers were investigated as the modified samples. 
A data set from reference and modified samples was obtained with NIR measurements. In this study, first Savitzky-Golay derivative with second polynomial order and window size of 15 was applied. It was compared with the usefulness of raw spectrum and also the combination of raw and first derivative spectrum. For the self-learning sector, an autoencoder neural network was employed. Within this thesis, it was shown that the trained autoencoder using first derivative spectra was capable to detect 5 % added water and 9 % cleaning solution in the milk. However, by using the combination spectra, the difference of 0.1 % in fat concentration was perfectly recognized. These two procedures were able to detect milks from different suppliers and difference of 10 °C in the measurement temperature.
Another part of this work was done using Raman spectroscopy. The aim of this part was to check if the previous result can be improved. In this step, the circulation method was again employed the same as in the previous part. However, because of the heat introduced to the sample by the laser using in Raman spectroscopy and the length of plastic tubes which can be affected by the temperature of the laboratory, the measurement temperature was kept at 10 °C. 1.5 % fat UHT milk was utilized as the reference sample. Milks with various changes such as different fat contents, various measurement temperatures and added water or cleaning solution were investigated as the modified samples. In this investigation, not only the autoencoder but also some chemometric models were utilized with the purpose of anomaly detection. Principal component analysis (PCA) was investigated to check if the various samples can becategorized separately. In addition, two chemometric modelling techniques such as principal component regression and Gaussian process regression were tested to check the ability for change detection. 
By using the results obtained by PCA, a sufficient categorization of various samples was not achieved. While the PCR did not present a promising prediction as the related R2 was 0.7, Gaussian process regression with R2 of 0.97 predicted the changes almost perfectly. The trained autoencoder and Gaussian process regression both were able to define 5 % water and cleaning solution, difference of 0.1 % fat content, and variation of 5 °C in the measurement temperature. 
In comparison between the autoencoder and Gaussian process regression, it should be mentioned that the Gaussian process regression was capable to determine more abnormal signals than the autoencoder, however, it must be trained with all the possible changes. In contrast, the autoencoder can be trained once just with reference signals and used in online monitoring properly. 
As the final part and to detect which type of anomalies happened during the milk processing, several classification approaches such as linear discriminant analysis, decision tree, support vector machine, and k nearest neighbour were utilized. While decision trees and linear discriminant analysis failed to effectively characterize the various types of anomalies, the k nearest neighbor and support vector machine presented promising results. The support vector machine presented an accuracy of 81.4 % for test set, while the k nearest neighbor showed an accuracy of 84.8 %. As a result, it is reasonable to assume that both algorithms are capable of classifying various groups of data accurately. It can help the milk business figure out whats going wrong during the processing of milk.
In general, Raman spectroscopy produced better findings than NIR spectroscopy, because the typical absorption bands of milk components in NIR spectrometers may be impacted by high water absorption combined with substantial light scattering by fat globules. Additionally, the autoencoder as self-learning system was capable of correctly detecting changes during milk processing, however, classification algorithms can aid in obtaining more details.
KW  - FT-Raman-Spektroskopie
KW  - Fourier-Spektroskopie
KW  - Chemometrie
CY  - Hohenheim
PB  - Kommunikations-, Informations- und Medienzentrum der Universität Hohenheim
AD  - Garbenstr. 15, 70593 Stuttgart
UR  - http://opus.uni-hohenheim.de/volltexte/2022/2009
ER  -