Existing domain generalization fault diagnosis methods achieve satisfactory interpolation performance but struggle with extrapolation owing to two fundamental limitations: insufficient source domain coverage and the inability to verify whether learned features represent causal fault characteristics or spurious correlations. To address these challenges, this study proposes a multi-perspective domain-invariant network (MPDIN) with energy–density-based data augmentation. MPDIN employs bootstrap aggregation to train multiple feature extractors on strategically defined domain subsets, establishing hierarchical domain invariance by enforcing subset-level invariance through triplet loss and inter-subset consistency via correlation alignment. This multi-perspective framework effectively suppresses subset-specific spurious correlations while preserving genuine fault characteristics. The energy–density-based augmentation leverages the -proportional relationship between rotational speed and vibration energy to generate realistic extrapolation data beyond source domain boundaries, utilizing raw short-time Fourier transform power spectrograms to preserve absolute energy information essential for physics-based scaling. Experimental validation across four diverse datasets demonstrated substantial improvements in challenging extrapolation scenarios, achieving gains of 19–47%, whereas conventional methods showed significant performance degradation. Manifold analysis confirmed continuity and complete target–source integration, validating the attainment of true domain-invariant learning. Although limitations exist in time-varying scenarios, the proposed methodology provides a principled framework for industrial deployment where targets frequently exceed training envelopes.