GTSFlow: Knowledge-guided spectral-topology flow matching for data-scarce fault diagnosis in wide-body mining dump trucks

Reliable fault diagnosis of wide-body mining dump trucks is constrained by the scarcity of labeled fault data and by strongly coupled, non-stationary responses recorded across distributed sensors. Existing time-series generators commonly learn statistical correlations in the time domain, but they do not explicitly retain amplitude–phase spectral characteristics or account for how inter-sensor coupling evolves along the generative path. To address these limitations, this study proposes GTSFlow, a knowledge-guided spectral-topology flow-matching framework for multi-sensor data augmentation. GTSFlow formulates conditional flow matching in the complex Short-Time Fourier Transform domain to retain amplitude and phase information throughout generation. It further constructs a Dynamic Spectral-Topology Graph (DSTG), in which spectral coherence represents sensor coupling, and introduces a topology alignment schedule that evolves from the independent noise prior toward the coupling structure estimated from measured data. Experiments using real-world operational data evaluate distributional and spectral fidelity, cross-sensor consistency, component effectiveness, computational cost, and downstream diagnostic utility. GTSFlow achieves the lowest MDD of 0.6095 and CFID of 0.0118 among the compared methods. When its generated samples are used for hydraulic oil leakage diagnosis, DLinear accuracy increases from 68.92% to 78.04%, an absolute gain of 9.12 percentage points. These findings indicate that jointly modeling spectral representation and inter-sensor topology can provide plausible synthetic data and improve diagnostic performance when labeled industrial fault data are limited.