To address the core challenges in traditional few-shot fault diagnosis—where the two-dimensional time-frequency representation paradigm destroys the inherent ”time-frequency-scale” three-dimensional coupling structure of vibration signals, leading to the loss of cross-scale fault information, insufficient feature discriminability, and blurred classification decision boundaries under few-shot and strong noise conditions—this paper proposes a semi-supervised fault diagnosis method based on 3D time-frequency representation and 3D residual networks. This method reconstructs one-dimensional vibration signals into a feature array that completely preserves the 3D physical structure through multi-scale continuous wavelet transform. It designs an adaptive scale gating mechanism to achieve adaptive screening of faultsensitive frequency bands and noise suppression. Furthermore, it builds a 3D residual network to directly extract deep structural features in the three-dimensional space, and integrates hypergraph topological structures to construct a semi-supervised learning framework based on curriculum consistency regularization. Comparative and ablation experimental results on standard public datasets show that this method effectively breaks through the dimensionality bottleneck of traditional 2D representation. The diagnostic accuracy can reach 98.9% in the 5-way 5-shot few-shot task, providing a new technical paradigm for intelligent mechanical fault diagnosis under scarce labeled data conditions.