Knowledge-Based Temporal GCN: A Spatial-Temporal Fault Diagnosis Method for Blast Furnace Ironmaking Process With Imbalanced Data

Xiong zhuo Zhu, Chunjie Yang, Siwei Lou, Yuelin Yang

Abstract: Fault diagnosis is essential to the safe operation of the blast furnace ironmaking process (BFIP). Currently, BFIP fault diagnosis still faces the following challenges: difficult-to-analyze spatial-temporal dependencies, difficult-to-use on-site knowledge, and imbalanced fault samples. In this case, a knowledge-based temporal graph convolution network (KB-TGCN) is proposed to address the above problems. First, the knowledge-based graph structure is built on the spatial and variable levels by utilizing the locations of sensors and the calculation relationships between variables. Subsequently, 1-D temporal information extractors (TIEs) are embedded in the nodes of GCN, and the TIEs can capture the mutilscale temporal information while maintaining the original spatial relationships. Therefore, KB-TGCN can fully explore the temporal and spatial information of BF. Additionally, to overcome the issue of data imbalance, the model is trained end to end by a focal loss (FL) function with an data-specific balance factor. Finally, the method is verified on the BF dataset, and the classification accuracy is higher than the baseline methods and the general spatial-temporal diagnosis method.