Multivariate Trend Feature Learning: Dual-Branch Graph Neural Networks for Fault Diagnosis in Process Industry

In modern process industries, fault diagnosis plays an indispensable role in maintaining safe operation. Despite the fact that various intelligence diagnosis models have been developed, few of them would notice multivariate trend information beneficial for identifying fault types. However, when malfunctions occur in actual plants, relevant variables exhibit not only certain outliers but also abnormal upward or downward evolution tendencies. Inspired by this observation, we propose dual-branch graph neural networks based on trend and numeric feature learning. In the first branch, a trend attention module is introduced to aggregate local change scales endowed with positional encodings from original time series. The acquired states are delivered to the corresponding graph nodes, further implementing the structured interactions and transfers for trend information. Meanwhile, in another branch, the multivariate numerical features are handled within the parallel graph space. Finally, a feature fusion layer is designed to achieve joint representations of node states for predicting classifications. With the complementary learning pattern defined over numeric and trend dimensions, our approach can fully capture potential and homologous process information, thereby tangibly improving the diagnostic performance. Two real-world datasets from the coal gasification and methanol production are used for application testing. The results show that the proposed framework outperforms conventional graph models driven by numerical features, demonstrating effectiveness in engineering practice.