基于多源数据融合与卷积长短期记忆神经网络的聚合物挤出过程熔体密度监测方法  

Melt Density Monitoring of Extruder Extrusion Process Based on Multi-source Data Fusion and Convolutional Long Short-term Memory Neural Network

作　　者：张彬彬 陈祝云 张飞 晋刚[1,2,3] ZHANG Binbin;CHEN Zhuyun;ZHANG Fei;JIN Gang(National Engineering Research Center of Novel Equipment for Polymer Processing,South China University of Technology,Guangzhou 510641,China;Key Laboratory of Polymer Processing Engineering of Ministry of Education,South China University of Technology,Guangzhou 510641,China;Guangdong Provincial Key Laboratory of Technique and Equipment for Macromolecular Advanced Manufacturing,South China University of Technology,Guangzhou 510641,China;School of Electromechanical Engineering,Guangdong University of Technology,Guangzhou 510006,China)

机构地区：[1]华南理工大学聚合物新型成型装备国家工程研究中心,广东广州510641 [2]华南理工大学聚合物成型加工工程教育部重点实验室,广东广州510641 [3]华南理工大学广东省高分子先进制造技术及装备重点实验室,广东广州510641 [4]广东工业大学机电工程学院,广东广州510006

出　　处：《工程科学与技术》2024年第6期54-62,共9页Advanced Engineering Sciences

基　　金：国家重点研发计划项目(2022YFC3901900);国家自然科学基金青年科学基金项目(52205101);广东省基础与应用基础研究基金项目(2021A1515110708,2023A1515240021)。

摘　　要：聚合物挤出过程中熔体密度是影响产品质量的关键因素。由于挤出加工过程的高温、高压复杂工况,寻求能准确、在线监测聚合物挤出过程中熔体密度的方法是一个具有挑战性的问题。尽管基于机器学习的质量监测方法提供了一种解决方案,但在聚合物挤出加工过程中,由于数据类型、工艺参数、操作环境等多变性因素的影响,传统的机器学习方法可能难以捕捉聚合物加工中不同输入参数和输出质量参数之间的复杂关系,使得监测任务难以获得理想的准确性。本文提出了一种基于多源数据融合与卷积长短期记忆神经网络(CNN–LSTM)的熔体密度监测方法,用于在线监测聚碳酸酯–丙烯腈–丁二烯–苯乙烯共聚物(PC/ABS)共混体系的熔体密度。首先,通过实时采集安装在挤出机模头处的近红外、拉曼及超声3种传感器数据,对3种传感数据进行预处理并融合后作为输入;然后,通过合理设计的网络结构,构建CNN–LSTM监测模型,利用CNN的特征提取能力与LSTM的预测能力,最终实现对聚合物共混过程中的熔体密度的实时监测。基于独立开发的多源传感数据实时采集装置获取的数据,利用所提方法对PC/ABS共混挤出过程的熔体密度进行实时监测,结果表明:本文方法能够准确监测聚合物熔体密度,其在测试集上的均方根误差和决定系数分别为0.975 5、0.006 3 g/cm3,比传统的卷积神经网络方法、长短期记忆网络方法、岭回归方法、偏最小二乘回归方法、多层感知机方法和支持向量机回归方法具有更高的预测精度;本文方法的10次输入平均预测时间为1.523 5 s,能够满足实际生产过程的实时监测。综上所述,所提出的基于多源数据融合与CNN–LSTM的熔体密度监测方法显著提高了聚合物挤出过程中熔体密度的实时监测精度,为挤出过程中聚合物的质量提供了可靠的技术支持。Objective This study addresses the challenging task of real-time monitoring of melt density during polymer extrusion,specifically for the polycarbonate-acrylonitrile-butadiene-styrene(PC/ABS)blend system,a critical parameter profoundly impacting the quality of the final product.Ensuring precise control over melt density is imperative for achieving desired product characteristics and maintaining process stability in polymer blending operations.Methods The research proposes a novel methodological framework that integrates multi-source data fusion with a convolutional long short-term memory(LSTM)neural network architecture to address this challenge.Using an independently developed multi-source sensory data acquisition device,three distinct sensor modalities,near-infrared,Raman,and ultrasound,are strategically positioned at the die head of the extruder to provide real-time data streams.These sensors capture vital insights into the dynamic changes occurring within the polymer melt during the extrusion process.The proposed model effectively learns the intricate mapping relationship between sensory data and melt density by amalgamating these multisource sensory inputs and using the feature extraction capabilities of convolutional neural networks and the temporal dependencies modeling capabilities of LSTM networks.Results and Discussions The application of the proposed method demonstrates significant efficiency in real-time monitoring of polymer melt density by monitoring the melt density during the PC/ABS blending extrusion process.Empirical evaluations reveal a root mean square error(RMSE)of 0.975 g/cm^(3) and a coefficient of determination(R2)value of 0.0063,underscoring the superior predictive accuracy of this approach compared to conventional methods.In addition,the proposed method exhibits robustness in handling the inherent complexities and variabilities in polymer extrusion processes,thus offering a reliable solution for ensuring product quality and process efficiency.The average prediction time for ten inputs is

关 键 词：聚合物挤出加工 熔体密度 多传感器数据融合 卷积长短期记忆神经网络 在线监测 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] TQ320.66[自动化与计算机技术—控制科学与工程]