Deep learning models face a trust deficit due to poor generalization and 'black-box' interpretability. Conventional transfer learning, reliant on statistical alignment, fails to guarantee physical plausibility. We propose a Physics-Constrained Transfer Learning (PCTL) framework based on the core insight that while raw signals vary, intrinsic physical fault patterns—like harmonic structures in the envelope spectrum—remain domain-invariant. Its key innovation is a 'diagnosis-verification-feedback' loop where an external, rule-based PCV module quantifies the consistency between a diagnosis and its physical evidence. This consistency score guides a confidence predictor, compelling the model's confidence to align with physical rationality. Extensive experiments show PCTL achieves superior accuracy and embeds reliable self-assessment, demonstrated by a strong correlation between predicted confidence and physical consistency. This research offers a new paradigm for developing intelligent diagnostic systems that are accurate, physically interpretable, and trustworthy.