A Novel Adaptive LSTM Model for Polyester Oligomer Density Prediction

In polyester fiber production process, oligomer density is the most important indicator which reflects the quality of the polyester fiber and the stability of the polyester polymerization process. Accurate prediction of oligomer density is essential for monitoring the subsequent series of chemical reactions. However, oligomer density is affected by numerous process factors. The use of raw high-dimensional process data for oligomer density prediction will leads to high model complexity and training difficulties. Therefore, this paper proposes a novel adaptive long-short term memory (LSTM) model based on attention mechanism to solve the above problem. First, a novel multi-head self-attention (MSA) mechanism is introduced to assign different weight coefficients to different feature variables, which can improve the ability to extract effective information and memory capability of the model. Then, concatenating a LSTM neural network to further learn the sequential features. Finally, a dense layer is utilized as a predictor to complete the prediction of key indicators. The proposed model named MSA-LSTM is validated by predicting the oligomer density of a polyester polymerization process.