In order to improve the level of power transformer winding condition monitoring, a winding deformation pattern recognition method based on sparse decomposition of frequency response curve was proposed. The core of this method is the matching pursuit algorithm in the signal sparse representation theory. Moreover, the over-complete dictionary of Gabor atoms was constructed and the simulation curves of frequency response were obtained by the finite element model of transformer. Based upon these, the sparse representation of the frequency response curves of winding under normal condition and deformation condition was carried out on the over-complete dictionary. Then, the short frequency Fourier transform and superposition of all corresponding matched Gabor atoms were carried out to obtain the equivalent time-frequency distribution of the normal curve and deformation curve. Finally the time-frequency distribution values of two different curves were subtracted in order to obtain the feature vector of pattern recognition, which also could be used as a criterion to reflect the degree of deformation of transformer frequency response curve. Finally, support vector machine (SVM) is used to realize the identification of the simulation curve of different winding deformation. The simulation results show that the method proposed in this paper has a high reliability, and is suitable for transformer winding deformation pattern recognition.