Faculty of information Technology, Beijing University of Technology
National Natural Science Foundation of China, Grant/Award Numbers: 61803005, 61640312, 61763037; Natural Science Foundation of Beijing Municipality, Grant/Award Numbers: 4192011, 4172007; Key R & D project of Shandong Province, Grant/Award Numbers: 2018CXGC0608; and the Beijing Municipal Commission of Education.
为考虑发酵过程的动态特征对阶段划分的影响，提高模型预测精度，本文提出一种基于注意力LSTM的多阶段发酵过程质量预测方法。首先，将原始三维数据沿批次展开，对每个时间片矩阵进行偏最小二乘(partial least squares，PLS)分析得到表征过程变量的得分矩阵和表征质量变量的得分矩阵，采用仿射传播(affinity propagation,AP)聚类算法将联合得分矩阵进行聚类，实现第1步划分；然后，采用encoder-decoder模型将表征过程动态性的动态特征提取出来，采用AP算法对其进行第2步划分；最后综合分析两步划分结果，将生产过程划分为不同的稳定阶段和过渡阶段，对划分后的各个阶段分别建立注意力LSTM集成质量预测模型；以青霉素发酵仿真数据和大肠杆菌生产数据进行实验测试，结果表明所提方法的可行性和有效性。
In order to consider the impact of dynamic features on stage division and improve the prediction accuracy, a quality prediction method based on attention LSTM is proposed. Firstly, the original data were unfolded along the batch direction. Partial least squares (PLS) analysis was performed on each time slice matrix to obtain the score matrix of process variables and quality variables. The joint score matrices were clustered by AP (affinity propagation) algorithm; then the encoder-decoder model was used to extract the dynamic characteristics of the process dynamics, and the AP algorithm was used for the second division. Finally, the production process was divided into different stable phases and transition phases through the comprehensive analysis of the two-step division results.The LSTM integrated quality prediction model was established in each stage after the division. The experimental data of penicillin fermentation simulation data and E. coli production data were tested and the results demonstrate the feasibility and effectiveness of the proposed method.