复杂流程工业过程的安全、平稳运行对于保证经济利益和社会效益具有重大的意义。数据驱动方法适用于复杂流程工业过程建模，并被广泛地应用于过程监控、诊断和质量预报等问题中。然而，传统数据驱动模型的物理意义不够明确，未能充分利用流程工业数据特点，导致数据中的大量宝贵信息未被挖掘，进而引发在实际应用中的一系列局限，制约了流程工业信息化与自动化水平的进一步提高。本文从流程工业数据的动态特性出发，提出了更加适用于工业过程建模的数据驱动方法，在此基础上对过程监控与诊断、控制性能评估以及软测量中的若干问题展开了研究。全文的主要内容与创新点如下：1.基于工业过程内在变化的缓慢特性，利用慢特征分析对过程变量间的稳态关系及动态关系同时进行描述，提出了全新的监控统计量，分别监控过程的稳态工况迁移及动态特性异常，为检测过程异常状况提供更全面的信息。该方法克服了静态监控模型未能监控动态特性的不足；克服了传统动态监控模型无法区分工况正常变化与动态异常的不足。2. 提出了基于SFA的控制性能监控方法，并结合贡献图技术对控制性能进行实时诊断，减轻了污染效应的影响，充分利用各类变量信息。针对过程时变特性，提出了高效的递推慢特征分析算法及相应的自适应监控策略；改进了模型更新停止准则，有利于监控系统自适应运行。3. 为了改进静态模型的局限性，分别从模型状态与参数的时序相关性入手，提出了两种新的动态软测量方法提出了概率慢特征回归模型，有效从时序数据中提取动态特征，实现多率数据融合；提出了辅助变量与主导变量采样时间不一致的建模方法。对动态偏最小二乘模型的参数光滑性进行改进，减轻了过拟合影响。4. 提出了一种新的动态非线性软测量方法，分别利用带有纯时延的一阶惯性环节和支持向量机描述过程动态特性及非线性。基于贝叶斯框架对模型参数进行迭代寻优，有效地减轻了非线性环节带来的过拟合影响，充分利用所有质量数据，具有良好的实用性。论文通过仿真案例及实际生产案例，对上述内容的有效性进行了验证。
Safe and steady operations of complex industrial processes are of significant importance to maximize economic benefits and social benefits. Data-driven methods are suitable for modeling complex industrial processes, and thus have been widely applied to areas such as process monitoring, fault diagnosis, and quality prediction. However, traditional models fail to make full use of unique characteristics of process data, with less interpretable physical meanings, thereby leaving a large quantity of information unused. This leads to a series of practical limitations, and prohibits the development of industrialization and informatization in the process industry. Based on the dynamic characteristics of industrial process data, data-driven methods are developed in this dissertation, which are better suited for industrial processes, then several problems are studied with respect to process monitoring, fault diagnosis, control performance assessment and soft sensing. The layout of this dissertation and its main contributions proceed as follows.Firstly, based on the slowly-varying nature of the essential variations of industrial processes, slow feature analysis (SFA) is introduced to simultaneously describe static and dynamic relationships among process variables. Then new monitoring statistics are put forward for concurrent monitoring of operating condition deviations and process dynamics anomalies, thereby furnishing comprehensive information about abnormal events of processes. The proposed method overcomes the deficiency of static models that dynamics of processes remains unmonitored, and enables an effective discrimination between nominal operating condition switches and temporal anomalies, which classical dynamic monitoring approaches fail to achieve.Secondly, a control performance monitoring approach is proposed on the basis of SFA, along with a real-time diagnosis method using contribution plots, thereby alleviating smearing effects and making full use of information that underlies different kinds of process variables. For time-varying processes, an effective recursive SFA algorithm is proposed, along with the associated adaptive monitoring scheme. In addition, the traditional stopping criterion for model updating gets improved, which is beneficial for long-run operations of monitoring systems.Thirdly, focusing on the deficiency of static quality prediction methods, temporal coherence is incorporated a priori into model states and parameters, respectively, yielding two new dynamic soft sensing methods. Probabilistic slow feature regression is proposed, which extracts useful dynamic features from time series data and synthesizes information of multi-rate data; meanwhile, an approach is proposed to tackle the case with non-aligned sampling instances of secondary variables and primary variables. The classical dynamic partial least squares model is enhanced in terms of temporal smoothness, and the overfitting phenomenon gets alleviated.Finally, a new dynamic nonlinear soft sensing approach is developed. Process dynamics and nonlinearity are described by, respectively, first-order systems with pure delays and support vector machines. Model parameters are iteratively optimized within the Bayesian framework, and the overfitting phenomenon induced by model nonlinearity is mitigated. All available quality data samples can be used for modeling, and thus the proposed method enjoys desirable practicability.Simulation cases and real industrial cases are adopted in this dissertation to testify the effectiveness of the proposed methods.
