Top-Blown Furnace Sensor Fault Diagnosis Based on Causal Graphs and Disentangled Learning

A top-blown nickel smelting system is equipped with a large number of sensors to ensure operational safety, but the complex coupling within the sensor network presents significant challenges for diagnostic tasks. In recent years, graph-based models have become increasingly important tools for fault diagnosis in complex systems. However, typical approaches represented by graph neural networks (GNNs) still have certain limitations: their heavy reliance on correlation analysis leads to poor interpretability, and they tend to learn spurious confounding features as “predictive shortcuts,” introducing potential confounders that significantly reduce the performance and generalization capability of the model. This article therefore proposes a causal path attention learning-based GNN (CPAL-GNN). By guiding the causal discovery algorithm with expert knowledge, deeper causal relationships can be constructed, and causal paths can be extracted, and an attention mechanism is used to generate causal and confounding features. Subsequently, disentangling is performed to separate these two types of features, and intervention is applied through backdoor adjustment to eliminate interference from confounders. This approach addresses the issues of GNNs relying on correlation analysis and using predictive shortcuts, thereby improving the diagnostic accuracy and interpretability of the model.