Quality index prediction of polypropylene based on improved Transformer framework

Le résumé fourni par la source

Polypropylene is widely used across various domains. Melt Flow Rate (MFR) is a crucial index for evaluating the quality of polypropylene products. Existing methods predominantly focus on high model accuracy while overlooking a fundamental issue: the actual production process involves multiple tasks and states, making it difficult for a single model to adapt to diverse production conditions. To address the challenge of performing robust and adaptive measurements due to property variations across different grades, a soft sensor based on improved Transformer framework with a mixture of experts mechanism and interval guidance (TMoEI) is proposed. First, the Transformer Encoder captures temporal features by simulating the differential positions of physical sensors, providing a robust infrastructure for data processing. Second, the mixture of experts (MoE) approach effectively handles the complexity introduced by varying MFR grades, optimizing expert model allocation through an intelligent gating mechanism. Additionally, a classification model is trained to predict the grade identification of polypropylene products. Inter-zonally standardized input data are utilized to conduct mixed-expert training based on these identifications. The MFR prediction results derived from the polypropylene granulation process indicate that the coefficient of determination (R2) reaches 0.9968, mean relative error (MRE) remains within \(\pm 2.5\%\), and fluctuation is reduced by more than 50% in low-grade interval compared to Transformer, thereby demonstrating the effectiveness and superiority of the proposed method.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.