Attention-Based Multi-Level Fusion for NIRS Modeling

Near-Infrared Spectroscopy (NIRS) can directly provide product quality information at the molecular level, such as concentration, viscosity, etc. However, NIRS is overly sensitive to environmental noise. Therefore, how to make full use of the information of process variables to enhance the robustness of NIR modeling is the key. For this purpose, this paper presents a multi-level fusion model based on Attention Mechanism (AM). First, two-branch Convolutional Neural Network (CNN) is introduced to extract the feature of NIRS. Then, a fusion block based on AM is designed, which consists a cross-attention operation, a residual connection, and a Multi-layer Perceptron (MLP). The residual connection weights and integrates features of NIRS and process variables according to the cross-attention distribution, while the MLP further combines the features in the feature dimension. Additionally, considering different information exchanged ways of fused features at each level of two-stream CNN, two different fusion schemes are designed namely central and shared strategies. Experiments on the 2,6-dimethylphenol distillation process datasets demonstrate that proposed method outperforms some existing methods.