ABSTRACT Accurate estimation of blend ratios in PTFE suspension processing is important for maintaining stable product quality. But the nonlinear and time‐changing behavior of real industrial mixing reduces the usefulness of traditional single‐modal soft sensing methods. This study introduces DynaGATNet, a lightweight dynamic graph‐based multimodal fusion network that uses image, light, and flow‐rate signals for blend ratio prediction. The model uses CNN and GRU encoders to extract features, and it builds dynamic modality graphs based on feature similarity. Then it applies graph attention to achieve adaptive cross‐modal fusion, so the model can capture time‐changing dependencies that static fusion methods cannot show. Experiments on an industrial dataset show that DynaGATNet reaches strong performance (MAE = 0.071, = 0.934, Accuracy = 0.938) while using only 3.6 million parameters and 2.1 ms inference time.
Chen et al. (Thu,) studied this question.