样本不平衡下基于图卷积网络的化工过程故障诊断

Graph convolutional network-based fault diagnosis of chemical process under sample imbalance

关键词:  化工过程  故障诊断  样本不平衡  图深度学习  混合边缘感知焦点损失函数

Key words:  chemical process  fault diagnosis  sample imbalance  graph deep learning  hybrid margin-aware focus loss

摘要:

为解决实际化工过程故障样本匮乏,现有故障诊断模型在数据分布不平衡下故障诊断准确率低的问题,提出一种基于代价敏感多感受野时空图注意力网络(cost sensitive multireceptive fields spatio-temporal graph attention network,CSMRFSTGAT)故障诊断模型。该模型通过最大信息系数(maximal information coefficient,C_MI)加权计算,将化工过程采集的相关变量数据转换为拓扑图数据,利用图卷积网络(graph convolution network,GCN)的故障诊断模型设计出了多感受野图卷积模块(multireceptive fields graph convolutional module, MRFGCM)和时空图注意力模块(space-time graph attention module,STGAM),然后提出了混合边缘感知焦点损失函数(hybrid margin-aware focus loss,L_HMF),用于对较难识别样本施加更多的惩罚。将所提模型应用于田纳西伊斯曼过程(Tennessee Eastman process,TEP)和三相流(three-phase flow,TPF)数据集中多个不平衡场景下评估其诊断性能。结果表明:所提模型在TPF数据集中的分类精确率和F1分数分别达到了91%和92%以上,同时在TEP数据集中的分类召回率和F1分数均突破了99%；相较于机器学习模型、深度学习模型以及图深度学习模型,所提模型能更加有效地识别故障。所提模型在处理数据不平衡问题上具有优异的泛化性能,能有效实现样本不平衡下化工过程故障诊断。

Abstract:

To solve the problem of low accuracy of existing fault diagnosis models under imbalanced data distribution caused by insufficiency of fault samples in practical chemical process, a fault diagnosis model based on cost sensitive multireceptive fields spatio-temporal graph attention network (CSMRFSTGAT) is proposed. This model converts the corresponding variable data collected from chemical process into topological graph data through maximum information coefficient (C_MI) weighted calculation. Using the fault diagnosis model of the graph convolutional network (GCN), multireceptive fields graph convolutional module (MRFGCM) and space-time graph attention module (STGAM) are designed. Then, a hybrid margin-aware focal loss function is proposed to impose more penalties on samples which are difficult to recognize. The proposed model is applied to evaluate its diagnostic performance in multiple imbalanced scenarios of the Tennessee Eastman process (TEP) and the three-phase flow (TPF) dataset. The results show that the proposed model achieves the classification precision and F1 score of more than 91% and 92% in the TPF dataset, and meanwhile the classification recall rate and F1 score in the TEP dataset both break through 99%, respectively; It can recognize faults more efficiently compared with the machine learning model, deep learning model and graph deep learning model. The proposed model has excellent generalization performance in dealing with the data imbalance problem, and can effectively realize chemical process fault diagnosis under sample imbalance.