Abstract To address the issue of low quality prediction accuracy in small‐batch batch processes, this paper proposes a batch process final quality prediction strategy based on multi‐stage multi‐resolution feature selection (MMFS) and dual attention bidirectional long short‐term memory (LSTM) transfer network (DA‐BLSTM‐TN). First, to reduce the redundancy of pseudo variables caused by directly expanding the input data in two dimensions, an MMFS method is proposed, which obtains optimized low‐resolution for each variable in different stages. Second, to improve the model's ability to extract quality‐related features, a dual attention bidirectional LSTM (DA‐BLSTM) network is proposed. This network designs an attention module at the input end of BLSTM to focus on the input features related to quality. Meanwhile, the historical and future dynamic information of process data is added through a bidirectional LSTM network to describe the dynamics of batch data more accurately. Moreover, an output attention module is developed to further performs attention weighting on the output features of BLSTM, so as to solve the problem of feature loss caused by using only the output feature of the last moment for quality prediction. Finally, a quality prediction model based on the DA‐BLSTM‐TN is constructed, transferring the modelling knowledge of similar old process to the new process to improve the accuracy of quality prediction for small‐batch batch process. The application in the penicillin fermentation process verifies the effectiveness of the proposed method.
Yao et al. (Mon,) studied this question.