In this paper, a novel method is proposed to detect demagnetization fault in surface mounted permanent magnet synchronous motors (SMPMSMs). Hence, a theoretical analysis is presented for feature extraction and pattern recognition via torque analysis. In this analysis, impacts of the saturation profile, stator slots and the fault severity are taken into account. So, using torque spectra, an efficient index is introduced for demagnetization fault diagnosis in SMPMSMs. Based on the extracted features and the recognized pattern, time delay embedding (TDE) approach as a competent data mining technique is utilized to process the developed torque. Then, a criterion function is proposed to detect demagnetization fault. Furthermore, this criterion is used to estimate fault severity and determine demagnetization percentage. The two-dimensional (2-D) time stepping finite element method (TSFEM) is employed to model the SMPMSM under different demagnetization levels. The 3-D TSFEM is used to prove the theoretical and 2-D TSFEM results.