Deep learning meets visualization: A novel method for particle size monitoring in fluidized bed coating

Abstract

The accurate monitoring of particle size distribution (PSD) is essential for ensuring the quality of oral solid dosage forms during fluidized bed coating. Here, a novel deep learning visualization framework was developed to predict and visualize PSD values of pellets based on near-infrared spectroscopy (NIRS). A multi-head self-attention convolutional neural network (MHSA-CNN) was designed to extract local spatial features as well as global contextual information from spectra. Bayesian optimization was employed to fine-tune the hyperparameters of the MHSA-CNN, thereby ensuring optimal model performance. Comparative analyses demonstrated that the proposed MHSA-CNN outperformed traditional CNN and partial least squares (PLS) methods in predicting PSD values, highlighting its robustness and accuracy. To further refine the network architecture, conventional method and uniform manifold approximation and projection (UMAP) were utilized to visualize the feature representations of the MHSA-CNN across different layers. The visualizations provided critical insights into the relationship between layer-wise feature transformations and PSD values prediction, facilitating iterative optimization of the MHSA-CNN structure by adjusting the number of layers. This systematic approach not only enhanced the predictive accuracy of the model but also provided a deeper understanding of the network’s inner workings.