Using Neural Networks to Obtain Indirect Information about the State Variables in an Alcoholic Fermentation Process

Applying AI Tools for Modeling, Predicting and Managing the White Wine Fermentation Process

This paper reveals two of the challenges faced by Romania and proposes a sustainable and simple solution for its wine industry. First, substantial areas with vineyards that may produce qualitative wine, and second, the very low digitalization rate of industrial sectors. More precisely, this work proposes a solution for digitalizing the fermentation process of white wine, allowing it to be adapted for other control techniques (i.e., knowledge-based systems, intelligent control). Our method consists of implementing a pre-trained multi-layer perceptron neural network, using genetic algorithms capable of predicting the concentration of alcohol and the amount of substrate at a certain point in time that starts from the initial configuration of the fermentation process. The purpose of predicting these process features is to obtain information about status variables so that the process can be automatically driven. The main advantage of our application is to help experts reduce the time needed for making the relevant measurements and to increase the lifecycles of sensors in bioreactors. After comprehensive simulations using experimental data obtained from previous fermentation processes, we concluded that a configuration that is close to the optimal one, for which the prediction accuracy is high, is a neural network (NN) having an input layer with neurons for temperature, time, initial substrate concentration, and the biomass concentration, a hidden layer with 10 neurons, and an output layer with 2 neurons representing the alcohol and substrate concentration, respectively. The best results were obtained with a pre-trained NN, using a genetic algorithm (GA) with a population of 50 individuals for 20 generations, a crossover probability of 0.9, and a probability of mutation of 0.5 that uniformly decreases depending on the generations, based on a beta coefficient of 0.3 and an elitist selection method. In the case of a data set with a larger number of variables, which also contains data regarding pH and CO2, the prediction accuracy is even higher, leading to the conclusion that a larger data set positively influences the performance of the neural network. Furthermore, methods based on artificial intelligence applications like neural networks, along with various heuristic optimization methods such as genetic algorithms, are essential if hardware sensors cannot be used, or if direct measurements cannot be made.