Bibliographic citations
Tirado, V., (2024). Monitoreo del estado de fermentación de café mediante espectroscopía de infrarrojo cercano con técnicas de aprendizaje automático y aprendizaje profundo [Universidad Nacional de Trujillo]. https://hdl.handle.net/20.500.14414/22213
Tirado, V., Monitoreo del estado de fermentación de café mediante espectroscopía de infrarrojo cercano con técnicas de aprendizaje automático y aprendizaje profundo []. PE: Universidad Nacional de Trujillo; 2024. https://hdl.handle.net/20.500.14414/22213
@mastersthesis{renati/1046263,
title = "Monitoreo del estado de fermentación de café mediante espectroscopía de infrarrojo cercano con técnicas de aprendizaje automático y aprendizaje profundo",
author = "Tirado Kulieva, Vicente Amirpasha",
publisher = "Universidad Nacional de Trujillo",
year = "2024"
}
Coffee fermentation is controlled visually and manually, which causes incomplete or excessive processes that lead to coffee with undesirable characteristics. pH and total soluble solids (TSS) are reliable indicators to determine the end of the process, but fast, accurate, and chemical-free techniques are needed. In the present work, the feasibility of near-infrared (NIR) spectroscopy coupled with machine and deep learning techniques to assess the progress of coffee fermentation by modeling pH and TSS was evaluated. Unfermented and fermented coffee beans (Coffea arabica L.) var. Typica, Caturra, and Catimor (4, 8, 12, 16, 20, and 24 h) were acquired from S´ıcchez in Peru. The grains were ground, spectral profiles were extracted in the range 1100 - 2100 nm, and pH and TSS were evaluated. Partial Least Squares Regression (PLSR) models were developed with all wavelengths and the effect of ten preprocessing techniques was evaluated. A K-Fold cross-validation strategy (K=5) was followed for training and validation of the models and after determining the best pretreatments for each parameter, PLSR, Random Forest (RF), Multilayer Perceptron (MLP), and Long-Short Term Memory (LSTM) models were implemented. The models were optimized by selecting the 10% most relevant features with the ReliefF algorithm, except for the LSTM model. The optimized MLP models showed the best performances with R 2 values greater than 0.88 and RPD greater than two with the best ability to predict pH, followed by °Brix, in Catimor coffee, followed by Caturra and Typica. The findings obtained suggest that NIR spectroscopy has the ability to predict the evolution of pH and TSS during coffee fermentation for monitoring and control purposes.
This item is licensed under a Creative Commons License