Deep learning autoencoder no desenvolvimento de soft-sensor para processos de fermentação alcoólica em batelada

Abstract

Alcoholic fermentation is a biochemical process in which simple sugars are converted into commercially interesting compounds, such as ethanol, for example. The Brazilian ethanol industry reached a production of 35.6 billion liters in the 2019/2020 harvest. The production process is influenced by various variables, such as the type of yeast, bacterial contamination, sugar concentration in the reaction medium, feeding time of the fermenters, temperature, pH, among others, and this makes the control and optimization of the process complex, especially when these variables must be measured directly by conventional physical sensors or laboratory analyses. Unlike conventional measurements, soft sensors use mathematical algorithms, artificial intelligence techniques, or statistical models to infer or predict these variables based on indirect information, such as available measurements, historical data, or correlations between variables that classical models cannot explain. Batch processes, unlike continuous processes, present a difficulty regarding the use of techniques for the development of empirical models for inference, which is how the time variable is considered and incorporated into the model. An Autoencoder (AE) is an unsupervised deep learning network composed of an encoder and a decoder that can be used to reduce data dimensionality and facilitate model training. It has been widely used as a classifier in fault detection and diagnosis; however, its use as a soft-sensor is still relatively new. Based on this, the work aimed to explore the use of AE, as well as some techniques already widely used in the development of soft sensors such as artificial neural networks (ANNs) and long short-term memory (LSTM), focusing on the issue of incorporating the time variable into the models. All three networks showed good results in predicting ethanol concentration, with R² values above 0,9 and errors below 2,5. However, the LSTM network stood out as the most effective, presenting better metrics for the evaluated variables (R² = 0.9973, RMSE = 0.4336 and MAE = 0.3055). The LSTM-Autoencoder model, on the other hand, proved to be a promising alternative, capable of maintaining good accuracy even when using a smaller amount of information (R² = 0.9551, RMSE = 1.7722 and MAE = 1.5142), making it relevant for applications where computational efficiency is a critical point.