Real-time yield prediction in microchannel gas-liquid sulfonation via augmented convolutional long short-term memory-based soft measurement

Abstract

Real-time monitoring of gas-liquid sulfonation in microchannel reactors is challenging due to complex internal spatiotemporal dynamics and limited data availability, despite the reactors’ excellent heat and mass transfer properties. Therefore, this study proposes a deep learning-based measurement method that directly extracts key spatiotemporal information from reaction image sequences within microchannels, enabling accurate prediction of the yield level of sodium α-olefin sulfonate products. The core of the framework is a convolutional long short-term memory network and combines a TimeDistributed module to efficiently capture and analyze dynamic visual features. To address the issue of data sparsity in experimental studies, we developed a novel frame sampling temporal image augmentation strategy that significantly improves the temporal learning efficiency of the model by mining microscopic dynamic changes under macroscopic stable conditions. On the experimental data set, the augmented convolutional long short-term memory network model achieved an average accuracy of up to 97.44%, outperforming the model without augmentation by 19.66% and a traditional convolutional neural network by 9.94%. These results demonstrate that the proposed method is a robust and effective tool for monitoring microchannel gas-liquid sulfonation, paving the way for intelligent, data-driven control of complex micro-chemical processes.