Reliable bearing fault diagnosis is important for the stable operation of rotating machinery under complex and varying conditions. Existing graph-based diagnostic methods commonly rely on either predefined fixed topologies or fully data-driven graph learning, making it difficult to balance structural reliability with state-dependent adaptability. To address this issue, this paper proposes a Prior-Guided Heterogeneous Dynamic Graph Convolutional Network (PHDGCN). PHDGCN constructs two relation-specific prior graphs from causality-consistent temporal dependencies and frequency-domain covariate correlations and uses them as structural anchors. A history-conditioned residual learning mechanism then refines the two priors according to the current local signal state while preserving their structural constraints. The refined temporal and covariate graphs are processed by independent graph convolutional branches, and their complementary representations are fused for fault classification. Experiments on two public bearing fault diagnosis datasets show that PHDGCN achieves accuracy/macro-F1 values of 93.75%/93.34% and 89.47%/87.08%, respectively, outperforming the compared conventional and graph-based methods. Ablation and graph-structure visualizations further show the contribution of prior-guided dynamic refinement and the relation-specific adjustment behavior of the two graph branches.