Degradation prediction for complex systems such as inertial measurement units (IMUs) is challenged by the need for nonlinear modeling, limited physical interpretability, and multi-source heterogeneous data fusion. To address these issues, a degradation prediction model integrating physics-informed neural ordinary differential equations (PINN-NODE) and heterogeneous graph neural networks (HGNN) was proposed. A parallel dual-branch architecture was used. In the temporal dimension, continuous-time dynamics under physical constraints were modeled by PINN-NODE. Irregularly sampled data were handled. In the spatial dimension, coupling relationships between sensors and environmental factors were captured by HGNN. Environmental stresses were reconstructed using dynamic edge weights. Spatiotemporal features were fused by a multilayer perceptron (MLP) to obtain an end-to-end predictor. On the IMU degradation dataset, RMSE is reduced by 56.43%, and MAE is reduced by 58.65%, relative to an LSTM baseline. The proposed model is shown to provide improved prediction accuracy with enhanced physical interpretability.