面向故障诊断的一类复杂动态过程数据建模

This project intends to set a research work on data-based modeling and analytics for fault diagnosis of a class of complex dynamical industrial processes. Firstly, data robust modeling methods for nonlinear dynamical processes under complex noisy environment are studied, based on which process monitoring and fault diagnosis approaches are then proposed. Secondly, the kernel method is introduced for nonlinear dynamical probabilistic modeling, while the hidden Markov model is used for switch modeling and monitoring of the process. Thirdly, new data modeling, fault detection and identification methods are developed for those dynamical processes whose state may change from one to another. Fourthly, the transition process between two operating modes of the process is particularly studied, including transition mode identification, nonlinear characteristic analysis, and joint fault monitoring. Finally, state order selection, parameter optimization and model performance evaluation for both dynamical statistical models and subspace data models are studied. The aim of this project is to develop a class of new method for modeling and analytics of complex dynamical processes. Achievements of this project will provide significant theoretical and application contributions for improving operating efficiency and product quality to typical industrial processes.

本项目针对一类复杂动态工业过程的故障诊断问题，拟开展基于生产运行数据的建模与分析方法研究。具体研究内容包括：复杂噪声环境下非线性动态过程的数据鲁棒建模、监测与故障诊断方法；基于核学习方法的非线性动态过程概率建模与故障诊断；面向复杂噪声动态过程的隐马尔科夫数据建模与状态监测技术研究；针对时变动态过程的自适应数据建模、故障检测与识别；多操作工况生产过程的工况切换和过渡过程模式识别、非线性特性分析以及联合故障监测方法研究；动态多变量统计模型与子空间数据模型的状态阶次选取、参数优化以及模型性能评价等。项目的研究目标是提出一套适合噪声环境下复杂动态过程故障诊断的数据建模与分析新方法。本项目的研究成果对提高流程工业生产过程运行效率和产品质量具有重要的理论意义和实际应用价值。

本项目深入研究复杂噪声、数据缺失、多工况、非线性、时变等环境下的动态过程数据建模与数据驱动故障诊断方法，取得了在故障诊断领域有影响力的研究成果，初步建立了一类复杂动态过程数据建模、故障诊断和在线监测的理论体系。主要研究进展有：揭示了闭环控制对数据驱动故障诊断的影响机制，提出了一种数学模型约束下的动态过程数据建模与数据驱动故障诊断新方法；提出了基于深度学习与流形学习的非线性动态过程数据建模与故障诊断方法；提出了工业过程关键变量的变分贝叶斯高斯混合回归软测量新方法；提出了基于自适应因子分析和轨迹分析的时变动态过程数据建模与故障检测方法；建立了多工况情况下的过渡过程数据模型、工况识别和故障检测算法；针对大规模厂级动态过程，提出了一种基于分布式贝叶斯网络和分层流形学习的大数据建模和监测方法。项目取得的部分理论研究成果在大型天然气制合成氨生产装置的运行状态监测与故障诊断实际应用，研发了天然气合成氨生产装置数据分析与故障诊断软件，取得一定经济效益。