数据驱动的故障诊断方法无需对象的数学模型，适用于复杂工业系统的故障诊断，对于提高系统安全性和可靠性以及保障工业产品质量具有重要意义。但是数据驱动的方法本质上是以相关性分析而非因果分析为基础，因此，可能会因为变量中存在的混杂因子而给出伪相关关联，这将进一步影响模型的稳定性和故障诊断系统的性能。考虑到有向无环图(Directed Acyclic Graphs, DAG)模型能够直观表示变量之间的因果关系，而DAG-NOTEARS提供了一种高效、准确的从数据中学习DAG图结构的方法，本文将采用DAG-NOTEARS方法建立DAG图模型，然后研究以此为基础的故障样本挖掘、故障检测和故障因果溯源方法。本文的主要贡献如下：(1)考虑到现有基于图模型的故障检测方法在建模时需要大量先验知识或基于数据的建模效率低的不足，通过以滑窗方式应用DAG-NOTEARS、并给出最优滑窗长度，提出基于最优滑窗的DAG-NOTEARS的建模方法；然后针对现有方法所采用的图相似度指标无法检测有向边方向变化的问题，提出基于DAG平均相似度的故障样本挖掘方法。基于TE过程仿真和带钢热连轧过程的实验结果表明，该方法提高了故障样本挖掘的效率和准确性。(2)针对(1)中提出的方法无法用于在线故障检测的问题，首先针对单模态场景，提出基于DAG拟合误差的残差设计方法，可用于在线故障检测；然后在此基础上，基于模态聚类和DAG建模，提出了基于DAG的多模态故障检测方法。基于TE过程和CSTH模型仿真数据的实验结果表明，该方法在不需要预先知道模态数量的前提下，可以准确的对多模态模态数据进行聚类，而且故障检测结果在不以牺牲漏报率为代价的前提下，显著降低了误报率。(3)针对基于贝叶斯网络的故障溯源方法会受到混杂因子和中介变量不利影响的问题，提出基于 算子的贝叶斯网络推理方法；然后进一步引入稳定学习以提高计算效率，从而最终提出基于DAG模型和因果分析的故障溯源方法。基于TE过程仿真和带钢热连轧过程的实验结果表明，该方法能有效去除混杂因子和中介变量的影响，实现故障原因的快速定位。
The data-driven fault diagnosis method does not require a mathematical model of monitored plants and therefore is applicable to fault diagnosis of complex industrial systems. It is therefore of great significance for improving system safety and reliability as well as guaranteeing industrial product quality. However, the data-driven method is essentially based on correlation analysis rather than causal analysis, and therefore, it may lead to spurious correlation due to the presence of confounders in the variables, which will further affect the stability of the model and the performance of the fault diagnosis system. Considering that the Directed Acyclic Graphs (DAG) model can represent the causal relationships between variables directly, and DAG-NOTEARS provides an efficient and accurate method to learn the graph structure of DAG from data, this article will use the DAG-NOTEARS method to build a DAG graph model, and then study faulty sample mining, fault detection, and causal analysis for fault based on it.The main contributions of the thesis are summarized as follows:(1) Considering the shortcomings of existing graph model-based fault detection methods including requiring a large amount of prior knowledge in modeling or low efficiency in data-driven modeling, an optimal sliding window-based DAG-NOTEARS modeling method is proposed by applying DAG-NOTEARS in a sliding window manner and providing the optimal sliding window length. Then, to deal with the problem that the graph similarity index used by existing methods cannot detect changes in the direction of directed edges, a faulty sample mining method based on DAG average similarity is proposed. The experimental results based on Tennessee Eastman (TE) process benchmark and hot strip mill process dataset show that the method improves the efficiency and accuracy of faulty sample mining.(2) To address the issue that the method proposed in (1) cannot be used for online fault detection, a residual design method based on DAG fitting error is first proposed for a single-modal scenario, which can be used for online fault detection. Then, on this basis, a DAG-based multimodal fault detection method is proposed based on modal clustering and DAG modeling. The experimental results based on TE process benchmark and CSTH model simulation dataset show that the method can correctly cluster multimodal modal data with an unknown number of modes, and the fault detection results significantly reduce the false alarm rate without sacrificing the false alarm rate.(3) Considering at the problem that the Bayesian network-based fault isolation method will be affected by confounder and mediation, a do operator-based Bayesian network inference method is proposed. Then, by further introducing stable learning to improve the computational efficiency, a cause analysis method for fault based on DAG model and causal analysis is finally proposed. The experimental results based on TE benchmark process simulation and hot strip mill process dataset show that the method can effectively remove the influence of confounder and mediation, and achieve rapid fault localization.
