Multi-resolution supervised attention LSTM for final quality prediction of batch process

Abstract

Quality prediction plays an important role in batch process monitoring and control. In order to improve the accuracy of batch process final product quality prediction, a multi-resolution supervised attention LSTM (MR-SALSTM) is proposed in this paper. First, a multi-resolution feature selection (MFS) method based on Grey Wolf optimization (GWO) is proposed. The low resolution of process variables in batch process is optimized by MFS to reduce the redundancy of process data in sampling time, and the adverse effects of irrelevant variables on quality prediction are eliminated by setting resolution thresholds. Second, in order to fully extract the nonlinear dynamic quality correlation features of batch process, a supervised attention LSTM network is proposed. The attention mechanism assigns different weights to each input feature to enhance the quality-related feature and suppress the quality-independent feature. At the same time, mass variables are input into LSTM cells to guide the learning process of nonlinear dynamic quality-related features. Finally, the predicted value of the final quality of the batch process is obtained by an additional fully connected layer. The effectiveness of the proposed method was verified by injection molding process and penicillin fermentation process.