Batch Plant Regressor Architecture for Sparse-Data Quality Control

The patent segments the batch process into multiple phases (e.g., filling phase, reaction phase, discharge phase) and trains separate neural network models for each phase. This segmentation allows the system to achieve accurate modeling comparable to whitebox models while reducing computational complexity by focusing each model on specific phase characteristics rather than modeling the entire complex process at once.

The patent implements dynamic adaptability by enabling the neural network model to be selectively retrained for specific phases when phase characteristics change, rather than requiring complete model retraining. The system can adapt to changing educt characteristics, process parameters, or product quality targets by updating only the affected phase models, maintaining flexibility while reducing computational overhead compared to conventional approaches.

The patent employs lightweight neural network models for each phase that can be independently trained and discarded/retrained as needed, replacing the need for complex hybrid models. These simpler phase-specific models achieve sufficient accuracy for their respective phases without requiring the integration of multiple model types, thereby reducing overall system complexity while maintaining adequate modeling precision.

The patent performs preliminary action by pre-training neural network models for each phase using available historical data, and then enabling rapid adaptation to new conditions by retraining only the affected phase models. This approach allows the system to maintain high prediction accuracy while significantly reducing the time required for model updates compared to complete retraining of conventional blackbox or hybrid models.