基于栈式自编码器的精对苯二甲酸生产装置质量相关过程分散监测

Aiming at the critical need for quality relevant process monitoring of INVISTA technology purified terephthalic acid (PTA) production plant and considering the characteristics such as large scale, nonlinear, massive historical data and so on, the basic theory and engineering application technology about the quality relevant process decentralized monitoring based on stacked auto-encoders will be researched. This study will try to solve the following key theories and technical difficulties. The theory and method of the adaptive stacked auto-encoders (ASAE), which is able to adaptively determine optimal network structure and obtain the suitable abstract features, will be researched and proposed; Based on data driven integrating with technology process analysis, the whole process decentralization, which is able to cluster monitoring variables with similarly varying characteristics and to highlight the local effects of all kinds of faults, will be proposed; Under the framework of decentralization strategy and based on ASAE, the accurate description of product quality information, the effective fusion of quality forecast information and monitoring model prediction result, timely and accurate warning or alarm of quality relevant faults and quality irrelevant faults, accurate fault traceability diagnosis and so on will be researched and solved. Finally, the basic theory and engineering application technology about the quality relevant process decentralized monitoring of large-scale chemical plant based on ASAE will be formed. The large-scale INVISTA technology PTA production plant will be employed to demonstrate the industrial application, and the sensitive and accurate quality relevant process decentralized monitoring system of PTA production plant will be built.

针对INVISTA工艺精对苯二甲酸（PTA）生产装置对质量相关过程监测技术的迫切需求、以及大规模、非线性、海量历史数据等特点，研究基于栈式自编码器的质量相关过程分散监测基础理论与工程应用技术。力求解决以下关键理论与技术难点：栈式自编码器自适应确定理论与方法（ASAE），实现最佳网络结构的确定和最适抽象特征的提取；有标签样本和海量无标签样本的充分利用；结合流程分析、基于数据驱动的全流程分散，实现变化特征相似的监测变量聚类，凸显各类故障的局部影响；以及在分散策略框架下，基于ASAE产品质量信息的精确描述、质量预报和监测模型预测的有效融合、质量相关故障和质量无关故障的及时准确预警或报警、准确的故障溯源诊断等。形成大规模化工装置基于ASAE的质量相关过程分散监测基础理论与工程应用技术，并在大规模INVISTA工艺PTA生产装置中应用示范，构建灵敏、准确的PTA生产装置质量相关过程分散监测系统。

精对苯二甲酸（PTA）企业迫切期望引入质量相关过程监测技术保证生产安全、产品质量稳定，从而提升生产技术水平。PTA生产装置是典型的大规模化工生产装置，鉴于深度学习技术在流程工业过程监测上存在巨大的潜力、以及分散监测能有效应对大型化工装置的大规模系统特性。本项目研究了基于栈式自编码器的PTA生产装置质量相关过程分散监测技术，解决了以下关键理论与技术难点：①提出了基于内积正则化策略栈式自编码器的过程状态监测、基于自适应参数栈式自编码器的过程状态监测、基于互信息和栈式稀疏自编码器的非线性动态过程状态监测、基于故障相关变量选择栈式稀疏自编码器的过程状态监测、以及基于局部和全局特征结构保持栈式稀疏自编码器的过程状态监测，形成了栈式自编码器最适隐层结构的自适应确定理论与相应的过程状态监测方法；②提出了基于深度学习技术的分散集成栈式自编码器过程状态监测、基于动静态特征分散融合的过程状态监测、混合主成分分析和栈式自编码器线性非线性特征分散融合的过程状态监测、以及基于互信息分散策略的动态过程状态监测，形成了大型化工装置分散监测基础理论和过程状态监测方法。③提出了基于栈式自编码器的质量相关特征提取与过程状态监测、基于自适应空间分解的质量相关过程状态监测、基于核空间中质量相关优化建模与过程状态监测、以及质量相关与过程相关的并行监测，形成了基于栈式自编码器的化工装置质量相关和无关故障监测理论与方法。最终构建了大规模化工装置基于栈式自编码器的质量相关过程分散监测基础理论与工程应用技术，并在大规模PTA生产装置、燃料乙醇生产装置、以及酚醛树脂生产装置上应用示范。相关成果发表学术论文64篇，SCI收录62篇、EI收录62篇，中科院一区13篇，三大化工期刊5篇；申请11项国家发明专利、3项国家软件著作权登记，获省部级一等奖1项；培养毕业6名博士和14名硕士。