A novel CNN-LSTM model with attention mechanism for online monitoring of moisture content in fluidized bed granulation process based on near-infrared spectroscopy

Abstract

In the fluidized bed granulation (FBG) process, accurate monitoring of granule moisture content is crucial for quality control. Near-infrared (NIR) spectroscopy is a commonly used method for detecting key attributes of granules in FBG processes. However, its spectral characteristics is susceptible to influences from particle size and other factors. There is a need to optimize preprocessing methods, spectral bands, and model types for the NIR models. To reduce the complexity in traditional modeling approaches, this study proposes a novel deep learning model with attention mechanisms for moisture content. The network architecture, named CNN-LSTM-Attention, integrates the spatial feature extraction capabilities of convolutional neural networks (CNN), the sequence-processing capabilities of long short-term memory (LSTM) networks, and the global correlation capturing capability of self-attention mechanisms. This model possesses both feature band optimization and sequence modeling capabilities, enabling more accurate processing of complex spectral sequence data. The results indicate that compared to traditional partial least squares (PLS) and support vector machine (SVM), the CNN-LSTM-Attention model requires no elaborate preprocessing for NIR spectra and can improve the predictive accuracy. The prediction performance of the proposed algorithm is verified using the NIR datasets. The calibration set achieved an R² of 0.9777 and a root mean square error (RMSE) of 0.19, while the validation set yielded an R² of 0.9689 and an RMSE of 0.20. The research findings provide theoretical guidance and technical support for the rapid and accurate detection of quality attributes in the FBG process, helping to improve quality control in the granulation process.