15 August 2026 in SSRN Electronic Journal
Multi-sensor vibration measurements provide complementary information for diagnosing complex faults in rotating machinery. However, changes in rotational speed and load can shift both single-channel feature distributions and inter-channel dependencies, limiting the transferability of fusion models across operating conditions. To address this problem, a Prior-Guided Topological Domain Adaptation Network (PTDAN) is proposed. Each synchronized multi-sensor vibration window is modeled as a sample-level graph with fixed sensor identities, while a shared multi-scale temporal encoder extracts channel-wise node features. Window-wise channel correlations are used as soft priors and fused with feature-driven multi-head attention to construct sparse dynamic sensor graphs. For cross-condition adaptation, PTDAN aligns graph-level fault features, global topology representations derived from adjacency information and graph-Laplacian spectra, and class-conditional topology prototypes estimated from reliable target pseudo-labels. A staged loss ramp-up strategy is adopted to reduce early negative transfer caused by inaccurate pseudo-labels. Experiments on the SEU gearbox dataset and a multi-condition rolling-bearing platform demonstrate average diagnostic accuracies of 97.74% and 95.45%, respectively. The method also achieves stable performance in cross-speed, cross-load, and combined speed-load transfer tasks. These results show that explicitly modeling and adapting multi-sensor relational structures improves the robustness and reliability of vibration-based fault diagnosis under varying operating conditions.