Abstract: To address the problems that engineering fault sample annotation relies on equipment disassembly and professional experience, and that existing graph-based diagnosis methods suffer from inaccurate sample correlation characterization, insufficient utilization of supervisory information, and fixed edge weights that are difficult to adapt to changes in node representations under extremely low label rates, an Adaptive Gaussian Propagation and Dynamic Graph Convolutional Network (AGP-DGCN) was proposed. In this method, an adaptive weighted graph was constructed based on the Gaussian similarity between spectral samples and node statistical thresholds, and weakly correlated edges were filtered out by a two-end consistency criterion. Pseudo-labels were generated for unlabeled nodes with effective responses through class-quality-balanced harmonic propagation. The latent-space similarity of existing edges was calculated using intermediate representations from graph convolution, fused with the initial edge weights, and combined with residual connections to complete classification. Independent experiments were conducted five times on the XJTU gearbox dataset and the Nanchang locomotive depot bearing dataset for validation. The results show that when the training set label rates are 0.5% and 2.0%, with only 20 and 42 labeled nodes, respectively, the average accuracy of the proposed method is 98.44% and 92.82%, and the standard deviation is 0.14% and 0.21%, respectively, which are 15.70% and 2.57% higher than those of the KNN-DGCN configuration using fixed K-nearest-neighbor graph construction. This method can reduce the interference of low-confidence cross-class connections and inconsistent neighborhood information, improve diagnostic stability under extremely low label rates, and provide a reference for few-label fault diagnosis of rotating machinery.