Sub-synchronous oscillations induced by the interaction between direct-drive wind turbines and the grid pose a serious threat to the safe and stable operation of the power grid. To rapidly identify the triggering unit, a localization method for sub-synchronous oscillation source based on short-time Fourier transform (STFT) images and transfer learning is proposed. Firstly, compressive sensing technology is employed to transform output data into observation signals, and then the STFT is performed on the observation signals to obtain the mapping image with oscillation characteristics, and the link between the mapping image and the oscillation source unit is constructed. Secondly, an adversarial transfer learning architecture is utilized in conjunction with the power system to achieve rapid generalization of unlabeled oscillation data in the target domain. Finally, the traditional transfer learning method is introduced for comparison, the results show that the proposed method performs better in terms of localization accuracy and efficiency, and has strong noise resistance.