EVALUATION AND SELECTION OF OPTIMAL DEEP LEARNING ARCHITECTURE FOR PREDICTING THE ENDPOINT IN HIGH SHEAR WET GRANULATION FOR ANTACID TABLET PRODUCTION.

Objective: The purpose of this research was to evaluate and select the best architecture among native convolutional neural network (CNN), MobileNetV2, ResNet50V2, and EfficientNetB0 for predicting the endpoint of the high shear wet granulation process, with accuracy as the main evaluation metric. Methods: The dataset was captured from an industrial camera using static image analysis and was manually labeled as “NOT READY” and “READY” according to the traditional endpoint method based on the mixer’s ampere point in the granulator. The dataset contained a total of 180 images, which were split between training and validation sets. Native CNN and TensorFlow Keras application programming interface (API) were utilized with MobileNetV2, EfficientNetB0, and ResNet50V2 as base feature encoders. Hyperparameters, such as final Fully Connected (FC) layer width, dropout rate, and learning rate, were optimized for binary classification using Keras hyper tuning. Results: The best was the native CNN, it was also the fastest among the three other models, taking only 20-30 ms per step for inference during runtime, though it requires 9000 ms time for training, the longest time among the models. It achieved an accuracy of 98%, and a validation accuracy of 97%. Conclusion: The system was able to determine when a wet granulation process has reached its endpoint based on live images from a camera after being trained on previously labeled data. The native CNN was the best model, offering the fastest runtime performance and the highest accuracy. [ABSTRACT FROM AUTHOR]