A semi-supervised graph learning approach based on dual-stream gated fusion for industry equipment fault diagnosis under extreme label scarcity

Abstract

Industrial equipment fault diagnosis remains challenging under variable-speed operating conditions and extreme label scarcity, where vibration signatures are easily affected by speed-induced spectral variations, operational noise, and amplitude outliers. Although graph neural networks can exploit structural relationships among vibration samples, their application to fault diagnosis is still limited by unreliable initial features and scarce labels, which may lead to distorted graph topology, error-prone message passing, and over-smoothed node representations. To address these issues, this paper proposes the dual-stream gated fusion network coupled with a label propagation system (DSGF-LPS), a semi-supervised graph learning framework integrating robust graph construction, label propagation, and dual-stream gated fusion. Specifically, Hanning-windowed fast Fourier transform and Spearman-rank-correlation-based graph construction are first used to obtain noise-resistant frequency-domain topology. Then, the LPS expands the limited ground-truth labels to high-confidence pseudo-labels, alleviating insufficient supervision. Finally, the DSGF-Net adaptively fuses local spatial attention captured by Graph Attention Network version 2 and multi-scale spectral semantics extracted by Chebyshev graph convolutions, while LayerScale and global residual connections are introduced to mitigate over-smoothing. Experiments on two rotating machinery datasets demonstrate that DSGF-LPS achieves accurate and stable diagnosis under highly non-stationary and label-scarce conditions, attaining over 97% accuracy with only two labeled samples per class.