炼胶过程的特征生成迁移建模与概率排胶研究

Rubber mixing process is the first and important production stage of the tire industry, which also consumes the main energy input to the integrated route of tire production. The Mooney viscosity is a key indicator related to the end-use properties in rubber mixing processes. To achieve a uniform product quality, variations in the Mooney viscosity should be reduced. However, in most current rubber/tire factories, the Mooney viscosity cannot be online measured in a direct manner. .This project aims to propose a novel modeling framework for online prediction of the Mooney viscosity in industrial rubber mixing processes. First, for describing the error-in-variable batch process data with highly incomplete labels, a semi-supervised deep modeling method with multi-stage feature extraction is developed. Additionally, a robust prediction method and a probabilistic inferential rubber-discharging strategy are proposed. For many rubber mixing recipes, there are only limited amounts of labeled data because of the low sampling rate. By exploration of the critical modes hidden in batches, an adaptive method, which can actively generate labeled pseudo data, is presented to handle this issue. Moreover, to overcome the large variations in rubber mixing processes, a quality-related feature transfer modeling method is developed. By efficient information fusion of the existing source models, the prediction performance of the target models can be improved. Finally, a unified framework, which integrates the feature-based generative transfer modeling method and the probabilistic inferential rubber-discharging strategy, is developed. This framework will be applied to complex rubber mixing processes for accurate online prediction and discharging of the multi-recipe products. The uniformity of the product quality will also be enhanced. .With this project, we want to enhance the application of advanced information modeling techniques in industrial rubber mixing processes, and improve the state of automation and intelligence of the tire industry in China.

炼胶是轮胎制造过程的首道重要且能耗较大的工序，其产品混炼胶质量对轮胎成品性能非常重要。然而，在当前工业炼胶过程中，衡量该质量指标的门尼粘度却无法在线测量，直接导致混炼胶控制水平不高。.本课题面向此工业间歇过程，针对半监督有噪数据，发展多阶段特征提取的深度建模和稳健预测方法，并形成概率推理的排胶策略。针对门尼粘度抽检造成标签数据少的现状，挖掘批次间的重要模式，提出有效标签数据主动生成的自适应方法。针对多配方和较大的工况波动，提出质量相关的特征迁移建模方法，快速融合有用的源域模型信息，以提升目标域模型的预测性能。最终，形成一套系统的、适合多配方混炼胶门尼粘度的生成迁移建模方法与概率推理排胶策略，在原材料和工况波动等复杂情况下，实现门尼粘度的较准确预测，提升混炼胶质量的均一性。.课题对促进先进建模方法在炼胶过程的应用，提升我国炼胶工业的信息化程度和自动化技术水平，具有重要的研究意义和应用价值。

课题针对工业炼胶、聚合等化工过程重要质量指标难以实时测量而导致的产品质量波动较大的关键问题，提出了适合半监督有噪工业数据的深度特征提取和稳健建模框架，发展了生成对抗学习和迁移学习相互融合的软测量迁移建模与性能提升方法，形成了一套适合多配方复杂工况聚合过程的生成迁移建模方法与概率推理策略，实现重要质量指标的较准确在线预测，提升了其质量控制的性能。.主要研究内容和重要结果包括：(1)提出了多工况过程的软测量迁移建模方法；(2)提出了针对具有多工况不均衡样本过程的虚拟标签数据有效扩增方法；(3)提出了融合过程非线性和动态信息的主动学习和建模方法；(4)提出了聚合物质量相关的本质特征提取与分析方法；(5)提出了针对有噪工业数据的稳健深度核学习半监督软测量建模方法。通过工业炼胶、聚合等实际过程的重要质量指标建模与在线预测验证了上述方法的有效性。.本课题成果包括正式发表了研究论文21篇，其中SCI期刊论文15篇，包括Journal of Process Control、Industrial & Engineering Chemistry Research和IEEE汇刊等领域内知名期刊。授权了国家发明专利6项，获得计算机软件著作权登记14项。培养了硕士11位，1位获得浙江省优秀硕士论文，2位获得浙江省优秀毕业研究生。课题内容涉及多个学科领域，较好实现了交叉融合。通过本项目的研究执行，加强了国内在深度学习、迁移学习、生成对抗学习等机器学习新方法在过程建模的应用，为国内炼胶、聚合等化工过程生产实施产品质量在线预测和控制，提供了应用基础和技术支撑，对于促进智能数据建模方法在过程工业领域的应用和工业数据智能的发展，将产生积极的现实意义。