Application of artificial neural network on the prediction of microbial population and species during spontaneous fermentation of garcinia mangostana pericarp / Mohd Fikri Hakim Abdullah and Mohamad Sufian So'aib

• Main Authors: Abdullah, Mohd Fikri Hakim (Author), So'aib, Mohamad Sufian (Author)
• Format: Book
• Published: 2020.
• Subjects: Conference or Workshop Item; PeerReviewed
• Online Access: Link Metadata
• Notes: https://ir.uitm.edu.my/id/eprint/82456/1/82456.pdf
• Keywords: Fermentation, Industrial

In this study, an artificial neural network (ANN) was used to predict microbial population dynamics and species during the spontaneous fermentation of Garcinia mangostana pericarp. The study was conducted by collecting the experimental data from analysis of fermented garcinia mangostana pericarp and train the data by using neural network in MATLAB system. The model was developed based on trial and error at different neural network architecture, transfer function, and training algorithm. The input parameter consists of days of fermentation (0-100 days) and volume of fermenters (5 and 50 liters). The data set were trained by the artificial neural network using hyperbolic tangent sigmoid (tansig) transfer function and Levenberg-Marquardt (trainlm) training algorithm. Based on the results, the best neural network architecture for prediction of the microbial population were 2-7-7-3 (bacteria) and 2-7-6-1 (yeast), while for the microbial species was 2-5-4. The correlation coefficient (R-value) for the training performance for prediction of bacteria and yeast population showed R-value were 0.99299 and 0.9703 respectively, while for the bacteria species was 0.94244. Performance of neural network design was evaluated based on mean square error (MSE) and relative error. The result shown the MSE for the training performance for prediction of microbial population were 0.009557 (bacteria) and 0.01358 (yeast), while for microbial species was 0.1077. The average relative error for microbial population for bacteria and yeast was evaluated to make sure the accuracy of the predicted data. The relative error means the percentage of incorrect predicted data. Hence, the least value of the average relative error will be good for the neural network model that indicate the accuracy between experimental data and predicted data.