A multimodal convolutional neural network trained by in-situ OCT characterization for dynamic structural prediction of nanofiltration fouling

Simulation and prediction of the dynamic and visual characterization of membrane fouling development remains a major challenge in membrane science and engineering. In this study, a dynamic image-based prediction model for nanofiltration fouling was developed using optical coherence tomography (OCT) fouling images. Non-Local Means (NLM) denoising method combined with the K-Means clustering algorithm was first established to achieve high-precision denoising and fouling layer extraction on OCT imaging. This approach clearly and concisely quantifies the instantaneous information of the fouling layer, providing high-quality data for model training. Accordingly, a convolutional neural network (CNN) based autoencoder model was developed by incorporating multimodal data fusion strategy that combined fouling image features under varied operational condition parameters. Through multimodal deep concatenation and a dynamic channel masking mechanism, the model effectively established a correlation between the nanofiltration fouling and its visualized morphological structure. The decoder subsequently generated multi-channel predictive images, enabling dynamic prediction of fouling morphology at varied filtration intervals. Experimental validation demonstrated that the proposed model exhibited excellent performance in capturing structural information and predicting dynamic evolution of membrane fouling, achieving high accuracy in terms of SSIM and MSE metrics. This study offers a theoretical and technical reference for intelligent dynamic monitoring and prediction of membrane fouling, and provides new insights into the optimization of operational parameters in membrane separation processes.