Deep learning-based models have been widely used in compound fault diagnosis of rotating machinery. A compound fault, distinct from a single fault, encompasses a variety of individual faults, suggesting a robust correlation among different compound faults. However, conventional data-driven methods often neglect this a priori knowledge, potentially impeding the performance of compound fault diagnosis. To counter this problem, our study endeavors to fully exploit the a priori knowledge of compound faults through label correlation modeling and knowledge transfer using a hierarchical neural network. Initially, a 2-D convolutional neural network (CNN) extracts features from short-time Fourier transform (STFT) time–frequency maps, which are subsequently processed using the multihead attention module to discern fault class features. Then label information is fed into the label correlation embedding (LCE) module constructed by static and dynamic graph CNNs (GCNNs), fully capturing the label relationships of different individual faults. Finally, the features that carry sufficient label correlation information are input into the classifier of sigmoid activation for decoupling compound fault diagnosis. Experimental validation on two separate datasets reveals that the proposed method surpasses existing methods in terms of exceptional classification accuracy, thus offering a novel paradigm for compound fault diagnosis in rotating machinery.