Brushless direct current (BLDC) motors have been widely used in industry and factory automations, and electric vehicles. Interturn short circuit fault is one of the dominated faults for a BLDC motor, and this fault affects precision control, induces noise and vibration, and even causes motor burn down and fires. Hence, diagnosis of interturn short circuit fault of BLDC motor is of significance. This paper proposes a method that combines of transfer learning and features fitting to realize accurate fault localization and evaluation. First, the three-phase current signals of the motor stator windings are synchronously sampled. The one-dimensional current signals are transformed to an image, and then a transfer learning-based convolutional neural networks model is trained for fault localization. When the fault phase has been localized, the sensitive features are extracted and selected from the corresponding phase current, and then features fitting method is designed to qualitative evaluate the fault levels. Experimental results indicate that the proposed method can localize the faults with accuracy of 100%, and the relative average error of fault quantitative assessment is 4.33%. The proposed method shows potential applications for accurate localization and evaluation of stator winding faults in permanent magnet motor systems.