过程监测技术是保障工业过程安全运行和高效生产的重要措施。由于原材料变化、设定点变动、启停机等，实际工业过程通常具有多工况特性。工况数量多、切换频繁、新工况不断出现等特点，与数据非平稳性和不确定性等耦合，给传统多工况监测方法带来巨大挑战。本文将多工况过程监测问题转化为连续工况监测问题，提出了具有持续学习能力的连续工况监测方法，并在大型火力发电机组制粉系统中得到应用验证。本文主要内容与学术贡献如下：（1）针对多工况平稳过程监测问题，提出了具有持续学习能力的主成分分析方法，适用于每个工况数据平稳情况。为减缓工况敏感参数的变化，该方法采用了弹性权重巩固方法离线估计参数重要性，通过提取新工况特征和巩固已学习工况的知识，不断更新单模型参数，实现已学习和未来相似工况的监测。此外， 针对上述方法缺乏可解释性问题，提出了具有持续学习能力的稀疏主成分分析方法，借助稀疏表示提高模型可解释性，通过突触智能方法在线估计参数重要性，并给出了故障可检测的充分条件。 （2）针对具有慢时变特性的多工况非平稳过程监测问题，提出了具有持续学习能力的稀疏动态主成分分析方法。该方法能提取数据中的动态和静态隐变量，挖掘数据的时序信息，提高模型可解释性和预测性能。针对突触智能存在估计偏差和初始化初值固定等问题，提出了能够准确估计参数重要性的改进突触智能方法。 （3）针对具有不确定性的多工况非平稳过程监测问题，提出了具有持续学习能力的概率慢特征分析方法。该方法能够对隐变量和不确定性进行显示建模，通过最大期望算法估计模型参数，不仅降低不确定性的影响，提高早期异常的监测灵敏性，还能区别部分正常工况变化和真实异常。 （4）针对一般的多工况非平稳过程监测问题，提出了自适应无监督监测方法，能够区别正常工况变化和真实异常。首先，提出了自适应协整分析方法，能够处理具有长期均衡关系的非平稳数据。其次，提出了具有持续学习能力的递归主成分分析方法，能够监测剩余非平稳信息，以此建立全面监测框架。采用递归核密度估计计算阈值，并结合先验信息构建统计量，提高工况辨识和异常检测准确率。 最后通过连续搅拌釜式加热器和浙能舟山电厂制粉系统验证了所提方法的有效性。

Process monitoring is effective and significant to guarantee the process operation safety and efficient production of industrial processes. Owing to the changes of raw materials and set points, startup and shutdown, practical industrial systems generally operate under multiple modes. Multimode processes furnish the features of the large number, frequently switching and continuous emergence of new modes and etc, which are also coupled with nonstationary characteristics of data and uncertainty. These characteristics would bring great challenges to traditional multimode process monitoring. This dissertation focuses on multimode process monitoring, which is transformed into monitoring successive modes. Then, several monitoring approaches with continual learning ability are proposed and applied to the pulverizing systems of large-scale thermal power generating units. The major contents and academic contributions of the dissertation are summarized as follows.(1) Aimed at multimode stationary process monitoring, a modified principal component analysis with continual learning ability is put forward, which is appropriate for stationary data in each mode. To slow down the changes of mode-sensitive parameters, elastic weight consolidation is adopted to measure the parameter importance offline. The model parameters are continually updated by extracting new features and consolidating the learned knowledge, which is able to monitor the learned and future similar modes. Besides, in order to overcome the lack of interpretability of the aforementioned proposed method, a sparse principal component analysis with continual learning ability is developed, where sparse representation is utilized to enhance model interpretability. The parameter importance is estimated online by synaptic intelligence. Moreover, the sufficient condition for fault detectability is provided to understand the proposed method comprehensively.(2) For multimode nonstationary processes with slow time-varying characteristics, a sparse dynamic inner principal component analysis with continual learning ability is proposed to monitor successive modes. Dynamic and static latent variables are extracted, where time series features are mined to improve the model interpretability and prediction performance. In addition, aimed at the estimation error and fixed initial parameters of traditional synaptic intelligence, a modified synaptic intelligence technique is developed to measure the parameter importance accurately.(3) For multimode nonstationary processes with uncertainty, a probabilistic slow feature analysis with continual learning ability is established, where the latent variables and uncertainties are modeled and the optimization issue is settled by expectation maximization. It can reduce the influence of uncertainty and enhance the detection sensitivity of incipient faults, and distinguish partial normal dynamic variations from real faults.(4) For the general multimode nonstationary processes, an adaptive unsupervised method is built and a comprehensive monitoring framework is constructed, which can distinguish a normal mode switching and real faults. First, adaptive cointegration analysis is developed to deal with nonstationary data with long-term equilibrium relationship. Then, a modified recursive principal component analysis with continual learning ability is proposed to extract meaningful features of the remaining nonstationary data. Recursive kernel density estimation is employed to calculate the thresholds adaptively and the test statistics are designed with prior knowledge, which is beneficial to enhancing the accuracy of mode identification and fault detection. The effectiveness of the proposed methods in this dissertation has been illustrated by a continuous stirred tank heater, and pulverizing systems which are from thermal power generating units at Zhejiang Provincial Energy Group, Zhoushan, China.