Abstract: The frequency stability of such signals in time-frequency domain was widely used to be as a feature to identify specific emitters, such as civil or military radio. But the acquisition environment and equipment noise restricted the improvement of the recognition rate even though the frequency estimation accuracy is high enough. Based on the analysis of the instantaneous frequency of the probability distribution function, this paper, combined with the cloud model method of classification and recognition, presented the improved phase fitting algorithm. In the first place of this paper, the data which has been preprocessed was used in the phase fitting procedure; the second step was to obtain short time frequency estimates, which was used to establish the normal cloud model; at the end of experiment, the preprocessed data was applied for classification. This method has improved the ambiguity of the threshold judgment and improved the embodiment of fuzziness and randomness. The experimental results of the simulation and practical signals show that the identification of the suggested technique can reach 97%, which is more accurate than conventional methods, even when at the expense of not using excessive additional time.