中国科学院自动化研究所机构知识库

随着工业体系的高速发展与科技水平的不断提高，现代智能制造工业系统已经基本实现超大规模化、高度自动化以及高度智能化，以满足日趋增长的工业需求。由于工业过程的复杂性，导致系统运行中往往不可避免地存在故障的发生。为了确保整个生产过程安全可靠的运行，需要及时检测出系统是否有故障发生，并确定故障传感器所在的位置区域，以辅助工作人员进行后续的故障清除工作。此外，为了得到稳定的产品输出，需要对质量相关的故障进行实时检测，以判断发生的故障是否会对产品质量造成影响。因此，工业过程的故障检测与隔离技术作为实现上述任务的有力手段，对于整个工业过程的安全、可靠与经济运行起到至关重要的作用。

由于复杂工业过程的机理模型难以获取，数据驱动的故障检测与隔离方法成为了近年来工业领域的研究热点。传统的数据驱动方法存在着非线性过程特征表示困难，在线检测实时性不佳，以及故障传感器定位不够精确的问题。对于质量相关故障检测，传统方法在过程变量与质量变量之间建立的关系模型不能较好地用于检测系统中是否发生了对产品质量产生影响的故障。鉴于此，本文以数据驱动方法为基础，分别从故障检测、质量相关故障检测、故障隔离等方面进行深入研究，旨在提高故障的检测率以及故障隔离的精确程度，并准确判断故障是否会降低产品质量。本文的主要研究内容及创新点可归纳如下：

1. 针对基于核方法的非线性过程故障检测任务中存在核参数难以恰当选择、在线检测实时性不佳的问题，提出了一种深度非负矩阵分解（DNMF）的非线性过程故障检测方法。首先，利用神经网络建立了DNMF模型，该模型利用输入数据自动学习出一个恰当的非线性映射，并将输入映射到适合非负矩阵分解的空间中，得到了更好的非线性特征表示；然后，基于贝叶斯推理的子空间融合，设计了用于全空间检测的统计量及阈值，从而实现故障检测策略。最后，选择一个数值仿真、田纳西-伊斯曼工业过程以及污水处理过程进行了实验，结果验证了该方法相比于核方法具有在线计算量小，检测实时性佳的优势。

2. 针对非线性过程质量相关故障检测中过程变量与质量变量之间难以建立相对准确的模型，故障与质量的相关性判断不准确等问题，提出了一种深度变体偏最小二乘（DVPLS）的质量相关故障检测方法。首先，将偏最小二乘模型改进为变体偏最小二乘模型，此模型可以利用神经网络建立过程变量与质量变量之间的关系；其次，利用深度自编码器搭建了非线性网络框架，并对变体偏最小二乘进行非线性化得到DVPLS，该方法等价于一个核参数可学习的核方法；然后，在DVPLS的损失函数中引入一个正则项，以增强系统特征对质量预测的稳定性；之后，利用奇异值分解和子空间融合技术，设计了适用于该方法的统计量和阈值，以及质量相关故障检测策略。最后，选择一个数值仿真、田纳西-伊斯曼工业过程以及污水处理过程进行了实验，结果验证了该方法相比于其它对比方法能够高效地检测系统中的故障，并能准确指出故障是否会对质量产生影响。

3. 针对现有k近邻的故障检测方法难以用于质量相关故障检测，且基于k近邻的隔离方法在假设不成立时故障变量定位效果不佳的问题，提出了一种核偏最小二乘与k近邻相结合的质量相关故障检测方法（KPLS-kNN）以及k近邻的改进变量贡献量的故障隔离方法（MVCkNN）。首先，提出的KPLS-kNN利用预测质量变量的近邻关系实现了质量相关的故障检测；其次，提出的MVCkNN利用累计相对贡献率将变量贡献量变为相对变量贡献量，来突显正常变量与故障变量在故障样本统计量贡献上的差异；然后，MVCkNN在不需要假设输入服从多元正态分布的情况下，提供了更为准确的故障变量阈值确定方法。最后，选择一个数值仿真、田纳西-伊斯曼工业过程以及污水处理过程进行实验，结果验证了KPLS-kNN相比于其它对比方法具有更好的质量相关故障检测能力，且提出的MVCkNN能够将故障变量定位在更准确的范围内。

With the rapid development of the industrial system and the continuous improvement of the science and technology, modern intelligent manufacturing systems have basically achieved super-large scale, high integration, high automatization and high intelligentization to meet the ever-increasing industrial demand. Because of the complexity of the industrial process, there exist some inevitable faults in the system. In order to ensure that the whole production processes have safe and reliable operational environment, it is necessary to detect whether a fault occurs and locate the position of failure sensors, which can assist the maintenance staff carry out the fault clearance. In addition, quality-related faults need to be detected in real time, so as to judge if product quality is affected by the faults happening in the system. Therefore, fault detection and isolation as powerful means to achieve the above tasks plays a crucial role in the safe, reliable and economic operation of the industrial processes.

Because the mechanism model of complex industrial process is hard to obtain, data-driven fault detection and isolation methods have become research hotspots in the industrial field in recent years. The traditional methods exist some problems, such as difficulty in nonlinear process feature representation, inability to guarantee the  performance of the real-time online detection, and inaccuracy of the range of fault sensor location. Moreover, the relationship model between process variables and quality variables established by traditional methods can not be well used for detecting quality-related faults. In view of the above problems, this paper conducts in-depth studies from the aspects of fault detection, quality-related fault detection and fault isolation respectively based on the data-driven methods, aiming at improving the detection rate of faults and the accuracy of fault isolation, as well as accurately judging whether to happen some faults that reduce product quality in the systems. The main research content and innovations of this dissertation can be summarized as follows:

1. In the current nonlinear process fault detection based on kernel method, the kernel parameters are difficult to choose properly, the real-time performance of online detection is not good enough, and the representation of the potential structure of the system is tough. Therefore, a nonlinear process fault detection method based on deep nonnegative matrix factorization (DNMF) is proposed to attempt to overcome the above problems. In this method, the neural network is adopted to establish the DNMF model, which can automatically learn an appropriate nonlinear mapping from the input and map input to a space that is suitable for nonnegative matrix factorization to obtain a better nonlinear feature representation of input. Then, based on the subspace fusion technique of Bayesian inference, the statistics and corresponding thresholds are designed, thereby proposing a whole-space fault detection strategy by utilizing the system features extracted from the proposed DNMF model. The experiments are carried out in a numerical simulation, the Tennessee Eastman (TE) process and the wastewater treatment process (WWTP). The results verify that the proposed method has advantages of less online computation and better real-time detection compared with kernel methods. 

2. In order to tackle the problems in quality-related fault detection for nonlinear processes that the relatively accurate model between process variables and quality variables is difficult to establish and the relation between the fault and quality is incorrectly determined, a quality-related fault detection method based on deep variant partial least squares (DVPLS) for nonlinear industrial processes is proposed to monitor product quality. First, the partial least squares model is modified to obtain a variant partial least squares (VPLS) model that can adopt neural networks to effectively establish the relationship between process variables and quality variables. Then, the deep autoencoder (DAE) is used to build a nonlinear framework which is proved to be equivalent to a kernel method with learnable kernel parameters. Using this framework, VPLS is transformed into its nonlinear version that is called DVPLS. Next, a regularization item is designed and introduced into the loss function of DVPLS, which enables to enhance the stability of system features to quality prediction. Finally, based on singular value decomposition and the subspace fusion technique of Bayesian inference, the statistics and corresponding thresholds are designed, thereby proposing a quality-related fault detection fault detection strategy by using the features extracted from the DVPLS model. The experiments are carried out in a numerical simulation, the TE process and the WWTP. The results verify that compared with other contrast methods, the proposed method can efficiently detect faults in the system and accurately point out whether the faults affect the product quality.

3. Aiming at the problem that the existing fault detection methods based on k-nearest neighbor rule (kNN) are hard to monitor product quality, and the fault isolation method based on kNN have wrong fault variable location when the hypothesis is not standing, a quality-related fault detection method based on kernel partial least squares and kNN (KPLS-kNN) and a modified variable contribution method based on kNN (MVCkNN) for fault isolation are proposed. KPLS-kNN method realizes quality-related fault detection task by utilizing the nearest neighbor relation of predicted quality variables. MVCkNN adopts accumulative relative contribution to modify variable contribution to relative variable contribution, in order to  reflect the difference between normal variables and fault variables in the contribution of fault sample statistics. Besides, it does not need to assume for MVCkNN that the input follows multivariate normal distribution in the setting of fault variable threshold, which extremely improves the accuracy of fault variable location. The experiments are carried out in a numerical simulation, the TE process and the WWTP. The results verify that compared with other contrast methods, KPLS-kNN has a better ability to detect quality-related faults, and the proposed MVCkNN can locate fault variables in a more accurate range.