Knowledge-Enhanced Spatial-Temporal Causal Analysis for Fault Detection of Nickel Top-Blown Furnace Sensors

The data-driven approach has garnered significant attention in the field of fault detection and diagnosis of blast furnace sensors. However, existing spatial-temporal prediction methods have lacked the modeling of causal relationships in the generation process of spatial-temporal data. This has resulted in the construction of biased spatial-temporal representations under the influence of spurious correlations, leading to decreased performance in downstream tasks and reduced model interpretability. To tackle this challenge, this article proposes a knowledge-enhanced approach for spatial-temporal causal diagnosis of nickel top-blown furnace (TBF) sensors. The approach incorporates domain expertise and causal discovery to establish qualitative causal relationships among the sensors in the form of causal graphs, thereby capturing the spatial relationships between them. In addition, a spatial-temporal causal transformer prediction model is developed to analyze the residuals between predicted and actual values, thereby achieving the objective of sensor fault detection. Experimental results demonstrated the accurate diagnosis of nickel TBF sensor faults using this approach. Analysis of the fault outcomes validated the comprehensibility of the model’s decision-making process in most cases, and its alignment with domain knowledge.