Permanent magnet synchronous motor (PMSM) is critical to industrial systems due to its high efficiency and precision. However, irreversible demagnetization faults induced by thermal or mechanical stresses compromise operational safety and reliability. While deep learning methods dominate current intelligent fault diagnosis, their inherent opacity and lack of physical interpretability limit deployment in safety-critical applications. This paper proposes the physics-informed reliable framework, a novel diagnostic paradigm that synergises electromagnetic physics with data-driven learning. The framework integrates physically interpretable features into its architecture by leveraging leakage flux characteristics, specifically fundamental frequency attenuation for uniform demagnetization and mechanical harmonic amplification for localized faults. The adaptive filter is able to extract fault-sensitive spectral components while maintaining alignment with electromagnetic principles. The performance of the framework was validated on a PMSM platform under 60 operational conditions (5 fault states, 6 speeds, 2 loads), achieving an average diagnostic accuracy of 99.5%. Spectral attention maps confirm consistency between model decisions and physical fault signatures, providing engineers with traceable evidence. This work bridges the gap between data-driven accuracy and industrial trustworthiness, offering a deployable solution for reliable motor health monitoring.