Application of Artificial Neural Networks for Predicting Cooking Dynamics in Industrial Sesame Seed Oil Extraction
Journal
Journal of Food Chemistry & Nanotechnology
ISSN
2471-4291
Date Issued
2024-10-01
Author(s)
Hamid Bakhshabadi
Alireza Ghodsvali
Abolfazl Bojmehrani
Mohammad Ganje
Toktam Mohammadi-Moghaddam
Elham Taghavi
Afsaneh Morshedi
Abstract
Sesame seeds are a significant source of vegetable oil and were among the earliest grains used for oil extraction. In this study, aimed at designing an industrial-scale process for extracting oil from sesame seeds, we investigated three cooking temperatures (75°C, 90°C, and 105°C) and three different moisture contents of the seeds leaving the cooking pot (4.5%, 5.5%, and 6.5%).
The study focused on several responses: the oil content of the pressed cake, the quantity of extracted oil, the protein and moisture contents of the resulting meal, and the percentage of insoluble fine particles in the extracted oil. To predict these responses, an artificial neural network (ANN) model was employed. Among the various backpropagation feedforward networks with different topologies studied, the configuration with 2 input nodes, 5 hidden nodes in one layer, and 5 output nodes was selected based on its high correlation coefficient (R² = 0.997) and low mean squared error (MSE = 0.0002).
The sigmoid hyperbolic tangent activation function was used, and the Levenberg-Marquardt learning algorithm with 1000 learning cycles was identified as the optimal neural model. The selected optimized models demonstrated high R² ≥ 0.97 during the evaluation of their results.
The study focused on several responses: the oil content of the pressed cake, the quantity of extracted oil, the protein and moisture contents of the resulting meal, and the percentage of insoluble fine particles in the extracted oil. To predict these responses, an artificial neural network (ANN) model was employed. Among the various backpropagation feedforward networks with different topologies studied, the configuration with 2 input nodes, 5 hidden nodes in one layer, and 5 output nodes was selected based on its high correlation coefficient (R² = 0.997) and low mean squared error (MSE = 0.0002).
The sigmoid hyperbolic tangent activation function was used, and the Levenberg-Marquardt learning algorithm with 1000 learning cycles was identified as the optimal neural model. The selected optimized models demonstrated high R² ≥ 0.97 during the evaluation of their results.