Abstract: With the development of radar technology, the diversity of radar system and the complexity of radar signal posed a severe challenge to the radar emitter signal recognition techniques. Cyclic bispectrum could be used as a feature to recognize the radar emitter signal for its excellent properties such as noise immune and carrying abundant information, but it has large amount of data. The main dimension reduction method lost most of the information. The symmetry and periodicity of cyclic bispectrum was proved, and a new method of feature extraction called partial axially integral cyclic bispectrum was proposed in this article. First， the new method calculated the cyclic bispectrum of the signal. Second, for each cyclic frequency, integrated cyclic bispectrum along the linear parallel to frequency axis on the plane constituted by the two frequency spectrum, and then transformed it into a vector. Finally the fisher discriminant rate (FDR) was used to select the features which have stronger ability to identify different signals. The new method not only could reduce storage effectively, but also utilized most of the useful information of cyclic bispectrum. Under simulation conditions, the performance of partial axially integral cyclic bispectrum was compared with the diagonal slice method. The result shows that the performance of new method is far superior to diagonal slice method.