Soft Sensor Development for Industrial Thickener using Neural Ordinary Differential Equation

The on-line measurement of underflow concentration,which is a crucial indicator for the thickening process, presents a significant challenge in mineral industry. This paper proposes a modeling approach employing a neural ordinary differential equation network with deep learning. The approach is designed to address the prediction challenges associated with underflow concentration in a thickener. Initially, we create a historical dataset using real data generated by a thickener operating under normal conditions at an industrial site. Subsequently, we train the neural ordinary differential equation network model for underflow concentration prediction. Finally, a subset of the dataset is selected to validate the accuracy of the predictions. The proposed approach achieves higher prediction accuracy compared to traditional deep learning methods.