As a key transmission component in the fields of wind power generation, rail transit, and aerospace, the reliability of permanent magnet synchronous motors (PMSMs) is crucial for the stability of the overall system. In actual operation, the motor is subjected to complex working conditions such as internal armature reaction, high temperatures and alternating loads,which are highly susceptible to irreversible demagnetization of permanent magnets.Therefore, it is significant to realize accurate diagnosis of demagnetization fault. In this article, a fault diagnosis method based on multi-modal data fusion is proposed to solve the coupling representation problem of loss-of-excitation fault in multi-physical signals.Based on the mechanism of electromechanical energy conversion, the relationship between electromagnetic torque harmonic and air gap magnetic field distortion caused by permanent magnet loss of excitation is explained, and the dynamic evolution law of vibration and current signal distortion is revealed. On this basis, a multi-modal feature extraction framework for vibration and current signals is constructed, which achieves in-depth mining and complementary enhancement of fault information by integrating spatial and temporal domain features. To improve the ability of fault identification and localization, a classification mechanism based on fusion features is designed, and the visual expression of diagnosis results was completed. Physical prototypes of motors with normal magnetic field, 10% uniform demagnetization, 20% uniform demagnetization, and local demagnetization are manufactured. Multi-modal data under normal and different demagnetization fault conditions are acquired by an experimental platform for verification. The results show that the diagnostic accuracy of the proposed method can reach 99.3%. Fault information can be collected conveniently and accurately via high-precision sensors without damaging the original structure of the permanent magnet motor. The effectiveness and engineering applicability of multi-modal data fusion for demagnetization fault diagnosis of permanent magnet synchronous motors are verified, showing promising application prospects.