Utilizing ResNet for enhanced quality prediction in PET production: an AI-driven approach

Abstract

To promote theoretical understanding for optimizing the entire process parameters (temperature, pressure, flow rate, etc.) and quality indicators (molar fraction, end-group concentration, and number-average molecular weight) in the industrial production of polyethylene terephthalate (PET), a dataset construction for production parameters and product quality indicators was accomplished in conjunction with industrial process simulation software. A complete deep learning workflow including data collection, dataset construction, model training, and evaluation was established. A prediction method for process-product quality of PET production based on the residual neural network (ResNet) network was proposed to reduce the complexity of quality control in polyester production. The results show that compared to traditional convolutional neural network (CNN), ResNet has higher accuracy (R 2 ≥ 0.9998) in predicting the PET production process and product quality. It can accurately establish the mapping relationship between production parameters and product quality indicators, providing theoretical guidance for intelligent production.