Using Stacked Auto-Encoder and Bi-Directional LSTM For Batch Process Quality Prediction

Abstract

Batch process quality prediction has broad application prospects in manufacturing and chemical industries. However, during the final quality prediction of a batch process, the final target values may be related to the whole process track of the batch reaction. Thus, the final quality prediction problem embraces complex high-dimensional input and simple low-dimensional output, which also means a serious size mismatch between input data and predictive values. Motivated by these difficulties, a hybrid prediction model is proposed, which combines the advantages of stacked auto-encoder (SAE) and bi-directional long short-term memory (BLSTM) for the final quality prediction of a batch process. The feature extraction ability of SAE is used to obtain the low-dimensional features of historical process data along the time direction. Then, the validity of the framework was verified by taking penicillin fermentation as an example.