化学工业在国民经济中占有重要的地位，化工安全是化工发展的必备条件。尽管现代化工过程已经可以通过控制系统实现高度的自动化，但当过程中出现较大的异常变化以至于超出系统的控制范围时，过程就会进入异常工况。此时需要操作人员进行详细的诊断。然而，人往往会受限于自身的知识、经验、情绪，对异常工况的诊断可能不准确，操作可能产生误动作。因此，有必要开发一套异常工况的透明诊断系统以进一步提升化工过程安全水平。异常工况的透明诊断是指对异常工况给出早期、全面、准确、可理解的诊断结果。具体来说有三点核心要求。第一，实现对异常工况的早期透明检测。第二，能高效地发现化工过程变量之间的因果关系，从而能以严密的推理逻辑进行诊断工作。第三，对异常变量、故障根原因、传播路径等信息给出高准确、低误报的诊断结果。本文针对当前化工过程异常工况诊断不够透明的问题，提出了以下方法。首先，针对化工过程的异常工况检测不够早期、透明的问题，本文提出了基于正交自注意变分时序自编码器的故障透明检测方法。该方法可充分提取化工数据中的非线性、时序性，并使用注意力机制识别异常变量。在连续搅拌加热器、田纳西-伊士曼过程中的案例研究结果表明该方法相比传统方法具有更高的故障检测率、更低的检测延迟，且能正确地指示异常变量，使得故障检测结果更加透明。其次，针对化工过程数据非线性、富含噪音、存在控制回路等特点，本文提出了基于因果门控时序Transformer的大数据因果发现方法。该方法通过因果门控时序Transformer模型发现变量间潜在因果关系，并通过重排特征重要性方法排除虚假因果关系。在模拟数据集和实际连续重整过程中的案例研究表明，该方法可以高效地从大规模化工过程的历史大数据中挖掘出变量之间的因果关系。随后，针对化工过程异常变量的识别准确率低、误报率高的问题，本文提出了基于因果-属性重构网络的异常变量识别方法。该方法利用变量间的因果关系信息避免了传统异常变量识别方法中的污染效应，显著地提升了异常变量识别的准确率、降低了误报率。在连续搅拌反应器上的案例研究证明了该方法的有效性。最后，本文综合上述方法，针对连续重整过程的关键设备提出了一套异常工况透明诊断方法，并在某石化企业连续重整装置案例中证明了方法的有效性。

The chemical industry plays an important role in the national economy, and chemical safety is a prerequisite for its development. Although modern chemical processes can achieve a high degree of automation through control systems, when significant abnormal changes occur during the process that exceed the system‘s control range, the process enters an abnormal situation. At this time, operators need to conduct detailed diagnosis. However, operators are often limited by their own knowledge, experience, and emotions, and the diagnosis of abnormal situations may not be accurate, and misoperations may occur. Therefore, it is necessary to develop a transparent diagnosis system for abnormal situations to further improve the safety level of chemical processes.Transparent diagnosis of abnormal situations refers to giving early, comprehensive, accurate and understandable diagnosis of abnormal situations. Specifically, there are three core requirements. First, early transparent detection of abnormal situations. Second, efficient discovery of causal relationships between variables in the chemical processes, allowing for rigorous reasoning logic in the diagnostic work. Third, providing high-accuracy, low false alarm diagnosis results for abnormal variables, fault root causes, propagation paths, and other information. This thesis addresses the lack of transparency in current abnormal situation diagnosis in chemical processes and proposes the following methods.First, to address the problem of insufficiently early and transparent detection of abnormal situations in chemical processes, this thesis proposes a transparent fault detection method based on orthogonal self-attentive variational temporal autoencoder. This method can fully extract the nonlinearity and temporal characteristics of chemical data, and use the attention mechanisms to identify abnormal variables. Case studies on a continuous stirring tank heater and the Tennessee Eastman process show that this method has higher fault detection rate and lower detection delay compared to traditional methods, and can correctly indicate abnormal variables, making fault detection results more transparent.Second, this thesis proposes a big data causal discovery method based on the causality-gated time series Transformer to address the nonlinearity, noise, and control loop characteristics of chemical process data. This method uses the causality-gated time series Transformer model to discover potential causal relationships between variables and eliminates spurious causal relationships through the permutation feature importance method. Case studies on simulated datasets and an actual continuous catalytic reforming process show that the proposed method can efficiently extract causal relationships between variables from historical data of large-scale chemical processes.Third, to address the problem of low accuracy and high false alarm rate of abnormal variables identification in chemical process, this thesis proposes an abnormal variable identification method based on causality-attribute reconstruction network. This method uses causal relationship information between variables to avoid the contamination effect in traditional abnormal variable identification methods, significantly improving the accuracy of abnormal variable identification and reducing the false alarm rate. Case studies on a continuous stirred tank reactor demonstrate the effectiveness of this method.Finally, the above methods are combined to form a systematic transparent diagnosis method for key equipment in the continuous catalytic reforming process. The effectiveness of the method is demonstrated in a case study of a real-world continuous catalytic reforming unit in a petrochemical company.