In industrial tofu production, soy protein gelation dictates final product quality, yet it is still monitored largely by operator judgement, leading to appreciable batch-to-batch inconsistency. Although ultrasonic techniques have shown considerable promise in tracking protein coagulation, most applications remain limited to static evaluation of current gelation state, which limits proactive control strategies. To overcome this gap, we propose a framework that integrates ultrasonic transmission sensing with multi-step time series forecasting. In particular, a triple-attention Seq2Seq (TA-Seq2Seq) model is developed, in which waveform attention identifies gelation-informative spectral regions, recency-biased temporal attention up-weights the most recent network development, and cross-attention to adapt information retrieval to each prediction horizon. Furthermore, a dual-probe through-transmission configuration with solid-state coupling pads was innovatively designed to mitigate the destabilising effect of high gelation temperatures on ultrasonic signal acquisition. To leverage ultrasonic information, full attenuation waveforms were collected at one-minute intervals over a 40-min gelation. Under Leave-One-Group-Out cross-validation, the experimental results demonstrate that the proposed model is capable of multi-step gelation degree forecasting across horizons from +1 to +5 min. For short-term prediction at the late gelation stage (t = 35 min), the model achieved a mean absolute percentage error as low as 1.02%, while for early-stage forecasting (t = 15 min), it still maintained an overall accuracy above 91.80%. Overall, by coupling ultrasonic signal trajectories with sequence modelling, this work advances soy protein gelation monitoring from static state estimation to multi-step forecasting, providing a quantitative basis for anticipatory endpoint determination and model-assisted process decision-making in tofu manufacturing.
更多
原文链接
分享
引用
