Machine Learning-Based Soft Sensor for a Sugar Factory’s Batch Crystallizer

In the contemporary industrial landscape, the utilization of real-time data for the surveillance and enhancement of operational processes stands as an imperative, contributing significantly to the refinement of operational efficiency and product quality across a multitude of sectors. This work presents the development of a machine-learning soft sensor utilizing Multivariate Linear Regression (MLR), Generalized Regression Neural Network (GRNN), Decision Tree, and Support Vector Regression (SVR) based on real historical data from a sugar factory. The soft sensor is designed to estimate the Brix index in the vacuum batch crystallizer. Various models have demonstrated robust performance in predicting Brix sensor values. The framework involves three key steps: data pre-processing, model construction employing selected algorithms, and the evaluation of well-performing models. While non-linear techniques, specifically Generalized Regression Neural Network (GRNN) and Decision Trees, exhibited superior performance in line with evaluation criteria, linear methods, such as Multivariate Linear Regression (MLR), closely matched the effectiveness of these advanced approaches