Dual-branch fusion framework with graph attention networks for compound fault diagnosis in complex machinery system

The present research proposes a generalized diagnostic framework for complex machinery that integrates domain knowledge with learned signal representations while exploiting the physical coupling between subsystems. By modelling multi-component systems as heterogeneous graphs, the framework employs a dual-branch architecture: Branch A extracts handcrafted features grounded in failure mechanics, while Branch B learns complementary latent embeddings via an unsupervised pre-trained 1-D convolutional autoencoder. To prevent geometric imbalance and ensure both branches contribute equally to diagnostic performance, Normalized Principal Component Analysis (nPCA) equalizes the feature spaces before a two-layer Graph Attention Network (GAT) performs topology-aware spatial fusion. Final health states are classified using a Platt-calibrated Support Vector Machine (SVM) with a One-Class extension, producing probabilistic labels alongside a risk index and a normalized entropy uncertainty index for robust open-set rejection. Validated on the PHM-Beijing 2024 Final Stage subway bogie drivetrain benchmark, the pipeline achieves 95.1%, 87.1%, and 82.4% accuracy across single-component, component-level compound, and system-level compound fault tiers, respectively. Crucially, the topology-aware GAT spatial fusion provides a unique + 12.4% point accuracy margin exclusively on system-level compound faults, successfully decoupling highly non-additive joint fault signatures where traditional independent ensemble methods collapse. Results confirm that the learned GAT attention weights align with physical coupling pathways such as the high-weight Motor-Gearbox connection and the uncertainty quantification effectively identifies 181 fault-affected samples from 252 unlabelled test cases under deliberate domain shift. Furthermore, cross-dataset verification conducted on the public Case Western Reserve University (CWRU) and University of Ottawa benchmarks demonstrates a superior 96.39% mean precision under cross-load transfer alongside highly stable uncertainty indicators under severe non-stationary speed dynamics. Operating under a severe dataset shift, the Platt-calibrated safety gate outputs a 86.5% low-confidence (LOW_CONF) triage flag distribution. This mechanism explicitly converts epistemic uncertainty into reliable human-in-the-loop work orders rather than issuing overconfident incorrect predictions, demonstrating significant practical engineering value for safety-critical industrial assets. This framework provides a monotonic severity profile and a reliable rejection mechanism for out-of-distribution conditions, serving as a blueprint for deployable diagnostic systems in multi-component rotating machinery.