Radar common intra-pulse characteristics are sensitive to signal-noise ratio, thus they are not adaptable to the complex radar systems. To solve this problem, a method was proposed which applied the multi-wavelet packets characteristics of bispectrum to sort unknown complicated radar signals under the large-scale signal-noise ratio condition with a high sorting rate. The bispectrum of received signals was extracted and predigested to two dimensions characteristic. Then, multi-wavelet packets were used to extract characteristics from two dimensions of the bispectrum and the best characteristics were selected as the sorting parameters. The best characteristics of eight classes radar emitter signals were extracted, then kernelized fuzzy c-means was used to cluster and sort signals. The experiment results demonstrate that the characteristics of eight typical radar emitter signals extracted by this method have good performance on noise-resistance and clustering with the large-scale signal-noise ratio.