异常监测技术对于确保工业过程的系统安全和生产效率至关重要。由于进料波动、工况变化、设备老化和未知扰动等因素，实际工业过程普遍具有非平稳特性，即测量数据的统计特性会随时间变化。非平稳特性给传统异常监测方法带来了巨大的挑战，因为传统方法难以对非平稳数据进行建模，而且异常工况容易被非平稳趋势所掩盖。本文围绕非平稳工业过程的异常监测问题开展研究，针对其中的复杂特性和难点问题提出了创新方法，并在大型火力发电机组中得到了应用验证。本文的主要内容与学术贡献如下：（1）针对具有多工况特性的非平稳过程异常监测问题，利用隐马尔可夫模型进行工况建模，提出了一步维特比算法进行实时工况辨识，准确率高于多种机器学习算法。之后，提出了一种性能驱动的主元选择方法进行故障检测，该方法弱化了故障可检测的充分条件，获得了更好的监测性能。（2）针对具有时序相关性的非平稳过程异常监测问题，提出了动态平稳子空间分析方法。该方法通过时移技术构建增广数据矩阵，以挖掘过程数据中蕴含的时序信息。然后将高维非平稳数据投影至一个低维平稳子空间，通过求解优化问题获得平稳成分。该方法可有效提取时序特征，提高对非平稳过程的监测性能。（3）针对具有不确定性的非平稳过程异常监测问题，提出了概率平稳子空间分析方法。该方法对平稳和非平稳成分同时建模，而且将实际过程变化与不确定性分离。利用期望最大化算法估计模型参数，并给出了参数更新的闭式表达式。该方法可以降低测量不确定性的影响，提高对早期小幅异常的监测性能。（4）针对具有非线性特性的非平稳过程异常监测问题，提出了核概率平稳子空间分析方法。该方法将非线性数据映射到一个高维特征空间，并在其中建立线性模型。利用核技巧估计模型参数，而无需知道非线性映射的表达式。该方法能够处理测量变量间的非线性关系，可有效监测非线性非平稳工业过程。（5）针对关键性能指标相关的非平稳过程异常监测问题，提出了输出相关的共同趋势分析方法，证明了它是偏最小二乘的推广形式。该方法建立了输入和输出变量之间的关联模型，可以有效检测出火力发电过程中的异常工况，而且能够判断异常是否影响锅炉热效率等关键性能指标。本文所提方法的有效性通过田纳西伊斯曼过程、连续搅拌釜式反应器等基准测试平台以及浙能舟山电厂1030MW超超临界火力发电机组得到了验证。
Anomaly monitoring is crucial to ensure the system safety and production efficiency of industrial processes. Due to many factors such as feed fluctuations, operating condition variations, equipment aging, and unknown disturbances, practical industrial processes generally show nonstationary characteristics, that is, the statistical properties of measurement data change with time. Nonstationary characteristics raise huge challenges to traditional approaches for anomaly monitoring, because it is difficult for traditional methods to model nonstationary data, and abnormal conditions are easily masked by nonstationary trends. This dissertation mainly focuses on the problem of anomaly monitoring for nonstationary industrial processes, in which several innovative methods are proposed for some complex characteristics and difficult problems, with applications to large-scale thermal power units. The main contents and contributions of the dissertation are summarized as follows.(1) For monitoring nonstationary processes with multimode characteristics, the hidden Markov model is used to model operating conditions. Then, the one-step Viterbi algorithm is developed for real-time mode identification, which is more accurate than some machine learning algorithms. After that, a performance-driven component selection method is proposed for fault detection, which weakens the sufficient condition for detectability and obtains better monitoring performance.(2) Dynamic stationary subspace analysis (DSSA) is proposed for monitoring nonstationary processes with temporal correlations, which constructs an augmented data matrix using the time shift technique to mine the temporal information contained in process data. High-dimensional nonstationary data are projected to a low-dimensional subspace, and the stationary components can be obtained by solving an optimization problem. DSSA can effectively extract temporal features of process data, then improve the monitoring performance for nonstationary processes.(3) Probabilistic stationary subspace analysis (PSSA) is developed for monitoring nonstationary industrial processes with uncertainty. PSSA models stationary and nonstationary components simultaneously, and separates actual process variations from the uncertainty. The expectation maximization algorithm is utilized to estimate model parameters, and the update expressions are derived in closed forms. PSSA can reduce the influence of measurement uncertainties, and improve the monitoring performance for incipient faults with a small amplitude.(4) Kernel PSSA (KPSSA) is proposed for monitoring nonstationary processes with nonlinear characteristics. KPSSA maps nonlinear data into a high-dimensional feature space, in which a linear model is built. The kernel trick is used to estimate model parameters, without knowing the explicit expression of nonlinear mapping. KPSSA can deal with the nonlinear relationship between measured variables, and effectively monitor nonlinear and nonstationary industrial processes.(5) Output-relevant common trend analysis (OCTA) is developed for key performance indicator-related nonstationary process monitoring, which is proved to be a generalized form of partial least squares. OCTA establishes an association model between input and output variables, which can not only detect abnormal conditions in thermal power generation processes, but also determine whether they affect key performance indicators such as the thermal efficiency.The effectiveness of monitoring methods proposed in this dissertation has been verified with benchmark processes such as the Tennessee Eastman process and the continuous stirred tank reactor, and 1030MW ultra-supercritical thermal power units at Zhejiang Provincial Energy Group, Zhoushan, China.
