Causal inference-based knowledge graph completion using multi-head attention for fault diagnosis

Knowledge graphs (KGs) play a pivotal role in organizing complex industrial data. KGs provide structured knowledge support for intelligent fault diagnosis. However, their effectiveness is often hindered by inherent data incompleteness and the scarcity of high-quality supervision signals. Existing KG completion (KGC) approaches, although effective in many general domains, are prone to “fault principle bias” and “attention bias” when used for fault diagnosis. They learn spurious correlations rather than true causal mechanisms. In this article, a new framework termed the causal KGC network with multi-head attention (CKGCNA) is proposed to address these challenges. First, a comprehensive fault diagnosis KG is constructed to characterize the relationships between fault symptoms, causes, and components. Second, a causal inference-based completion model is developed. This adds both a multi-head self-attention mechanism that captures global structural information and an expectation module that performs front- and back-door causal adjustments. This dual-adjustment mechanism effectively mitigates the confounding effects of spurious correlations. Finally, the proposed method is used to analyze real-world, automotive fault diagnosis data. The results show that the proposed method significantly outperforms various state-of-the-art baselines in terms of link prediction tasks, confirming its superior ability to capture latent causal relationships and improve the completeness of fault knowledge bases.