Kolmogorov–Arnold and deep learning networks for industrial explainable product quality prediction

In the chemical process industry, the frequent measurement of quality variables is not always a viable path to take. It can be due to the measurement being time- or cost-consuming, leading to a sampling time of hours or even a day. As such, soft sensors were developed as a solution to this problem, providing a model trained on the available output measurements and predicting its value between samplings. This paper presents the process of soft sensor development for an industrial case study and the application of recently developed Kolmogorov–Arnold Networks (KAN) as a possible soft sensor model, where also the inherent interpretability of KAN through symbolic regression is analyzed. The soft-sensor development and interpretability analysis is performed on a benchmark dataset (steam turbine dataset), and on a real industrial case study (aniline synthesis plant). The prediction performance of KAN is compared to different complexity ML models, and in addition, the models were compared on the basis of their interpretability and explainability using Shapley values. The results show that the inherent interpretability of KAN models using symbolic regression is too complicated on simple tasks, thus losing its ability to explain the model estimations, though its model performance is similar to the classic feedforward neural networks.