Complex correlations among health indicators (HIs) critically impact prediction accuracy, yet existing State-of-Health (SOH) estimation methods struggle to simultaneously model their time-varying coupling relationships and cross-feature dependencies. To address this, this paper proposes an Explicit–Implicit Graph Fusion based Spatiotemporal Graph Convolution Network (EIGF-STGCN). Specifically, regularized graph structure learning is employed to generate sparse implicit graphs for mining latent correlations. These are dynamically fused with explicit mechanism-based prior graphs via self-attention to construct a structure adaptable to degradation evolution. Subsequently, a GCN-TCN model jointly captures spatial dependencies and temporal degradation patterns. Experimental results on NASA and CALCE datasets demonstrate that EIGF-STGCN outperforms LSTM, Transformer, and GCN-TCN. On the NASA dataset, the average MAE and RMSE are reduced by 32.67% and 36.78%, respectively, and by 18.10% and 24.39% on the CALCE dataset. Furthermore, interpretability analysis confirms that the implicit graph uncovers cross-group couplings missed by explicit mechanisms, while the dynamic fusion accurately characterizes time-varying dependencies throughout the lifecycle.