一种新颖的深度因果图建模及其故障诊断方法

为了实现复杂工业过程故障检测和诊断一体化建模, 提出了一种新颖的深度因果图建模方法. 首先, 利用循环神经网络建立深度因果图模型, 将Group Lasso稀疏惩罚项引入到模型训练中, 自动地检测过程变量间的因果关系. 其次, 利用模型学习到的条件概率预测模型对每个变量建立监测指标, 并融合得到综合指标进行整体工业过程故障检测. 一旦检测到故障, 对故障样本构建变量贡献度指标, 隔离故障相关变量, 并通过深度因果图模型的局部因果有向图诊断故障根源, 辨识故障传播路径. 最后, 通过田纳西?伊斯曼过程进行仿真验证, 实验结果验证了所提方法的有效性.

A Novel Method for Deep Causality Graph Modeling and Fault Diagnosis

To achieve integrative modeling of fault detection and diagnosis for complex industrial process, this paper proposes a novel deep causality graph modeling method. The deep causality graph model is constructed based on recurrent neural network. Group Lasso sparse penalty term is introduced into model training to detect the causality among process variables automatically. Then the conditional probability prediction models learned from deep causality graph model are used to establish single variable monitoring index for each process variable. An integrated global monitoring index is obtained for fault detection in the whole industrial process. Once fault is detected, the variable contribution indexes are built to isolate fault-related variables, and the local causality directed graph network built by fault-related variables is used for root cause diagnosis and propagation pathway identification. Finally, the proposed framework is simulated by the Tennessee Eastman benchmark process data. The experimental results verify the effectiveness of the proposed method.