Faced with the limited labeled sample problem in practical engineering, particularly in extreme labeling scenarios where only one labeled sample is available for each fault type, the existing semi-supervised diagnosis methods suffer from a significant deficiency in the fault identification ability. To address this issue, a novel semi-supervised fault diagnosis method based on the decoupled feature pseudo-label propagation (DFPP) algorithm is proposed. Firstly, the locally selective combination in the parallel outlier ensembles (LSCP) method is introduced to separate fault samples. Subsequently, the DFPP method is proposed. In DFPP, the adversarial decoupled auto-encoder (ADAE) is applied to extract the enhanced fault features, and the incorporation of fault feature dimension reduction, pseudo-centroid calibration of feature distribution, and distance measurement are adopted to efficiently achieve pseudo-label propagation in situations. Finally, a fault classifier is trained by using pseudo-labeled fault samples, and the combination of anomaly detection ensures accurate fault diagnosis with high precision. Experimental results conducted on two datasets of rotating components demonstrate that the proposed method can achieve average diagnostic accuracies exceeding 97% and 90% in the same working condition and cross working condition with extremely limited labeled samples, respectively, which is significantly superior to the comparison methods.