Developing an accurate and reliable anomaly detection model is of great significance for safe operation in the process industry. To minimize false positives, it is crucial to accurately model the intricate topological and nonlinear connections among variables. In this study, a process anomaly detection method based on graph-guided masked autoencoder (GGMAE) is proposed by introducing the concept of the graph to the process industry. GGMAE first constructs a topology graph according to the process flowchart. Then, the patch and mask mechanism forces GGMAE to learn the process intrinsic information to reconstruct the input according to the variable topological relationship and temporal characteristics. Additionally, Kullback-Leibler divergence is used as the loss to ensure that the distribution of input and output is consistent. Experiments on two publicly available anomaly detection benchmarks demonstrate the superiority of GGMAE over existing methods. Visualization of the results demonstrates the physically compliant reconstruction logic.

中文翻译：

---

用于过程异常检测的图形引导屏蔽自动编码器


开发准确可靠的异常检测模型对于流程工业的安全运行具有重要意义。为了最大限度地减少误报，准确建模变量之间复杂的拓扑和非线性连接至关重要。本研究通过将图的概念引入流程工业，提出了一种基于图引导掩码自动编码器（GGMAE）的过程异常检测方法。 GGMAE首先根据流程流程图构建拓扑图。然后，补丁和掩码机制迫使GGMAE学习过程内在信息，以根据可变拓扑关系和时间特征重建输入。另外，使用Kullback-Leibler散度作为损失，以确保输入和输出的分布一致。对两个公开可用的异常检测基准的实验证明了 GGMAE 相对于现有方法的优越性。结果的可视化展示了物理上兼容的重建逻辑。