Decanter centrifuges are well-established devices used for continuous solid-liquid separation in numerous industries. The separation performance is based on centrifugal force, which, due to density differences between the solid and liquid phases, enables sedimentation and subsequent discharge of the solids. Despite significant advances in process modeling, uncertainties remain in predicting operational behavior, particularly under varying material properties and transient process conditions.

This work presents a novel approach that utilizes continual learning methods to adaptively improve process models of decanter centrifuges during ongoing operation. Online analytical data is continuously collected and used to dynamically identify and correct model uncertainties. In contrast to classical static models, this approach enables continuous adaptation to changing process conditions and material properties without the need for extensive offline experiments or pilot tests.

A central component is the combination of physics-based process models and data-driven learning methods. While established models describe fundamental mechanisms such as sedimentation, sediment compaction, and solids transport, remaining model deviations are compensated for by learning algorithms. These algorithms update themselves incrementally based on new measurement data, thereby gradually improving the quality of predictions.

The results show that ...