Abstract
This study proposed a novel Bayesian Network model integrating Graph Attention mechanism and Adaptive Prior algorithm, termed GAT-BN, to address the challenges of data sparsity and imbalanced fault distribution in root cause analysis (RCA) of aircraft engines. The model incorporated domain-specific hierarchical constraints during the structure learning phase, ensuring that the derived network topology aligns with the physical fault propagation logic. For parameter learning, association rules mined from global data serve as robust prior knowledge. A hierarchical attention mechanism was designed to adaptively calibrate the prior strength. This mechanism effectively compensated for the scarcity of deep fault data. Moreover, the embedded GAT module autonomously learns node criticality scores, generating a neural-based prior distribution that enables the model to focus on low-frequency yet high-consequence root causes while mitigating the interference of high-frequency but low-criticality faults. Extensive evaluations on a dataset of 634 engine fault records, using Top-K hit rate and mean absolute error as metrics, demonstrate that GAT-BN performs better than traditional Bayesian Networks and other benchmark models under the experimental settings. It exhibits exceptional robustness under data-sparse scenarios. The proposed framework presents a new paradigm of structure-knowledge-data collaborative driving for intelligent fault diagnosis in complex industrial systems.