AI-driven soft-sensor for stirred media mills for real-time process control

Real-time monitoring of particle size distribution (PSD) and viscosity remains a challenge in wet grinding, particularly due to their complex interdependence. This study presents a sequential neural network architecture that predicts both parameters via a soft-sensor approach, eliminating the need for offline sampling and dilution. The approach predicts the PSD from ultrasonic extinction measurements and uses these predictions to estimate suspension viscosity. Notably, the system covers a range of particle sizes and viscosity values, which complicates accurate real-time measurements and modelling. Experimental validations are carried out on a laboratory-scale stirred media mill processing Al₂O₃ suspensions across operating conditions, with parameters including tip speeds (6–12 m/s), bead sizes (300–700 μm), and solids mass concentrations (10–30 %). Using data augmentation to enhance model robustness, the architecture achieves good accuracy (R² > 0.85) for PSD predictions and 0.82 for viscosity predictions. The model successfully captures process dynamics, including transition points where particle size reduction below 400 nm triggers significant rheological changes. A sensitivity analysis identifies mill tip speed and grinding bead diameter as the most influential parameters affecting prediction accuracy. This sequential approach enables continuous real-time process monitoring and lays the foundation for advanced control strategies in grinding operations.