In order to solve the shortcomings of signal separation methods based on the dictionary learning in phase information loss and cross representation, a signal separation algorithm based on the discriminative amplitude-phase dictionary learning was proposed. In discriminative amplitude-phase dictionary learning method, a model of amplitude-phase dictionary was proposed to solve the problem of phase information loss. Meanwhile, based on the idea of discriminative dictionary learning, a penalty term of cross representation was added into the object function of dictionary learning to solve the problem of cross representation, which happens to the mixed signal projected in joint dictionary. Experiment results show that the amplitude and phase information of low probability of intercept signals can be fully represented by amplitude-phase dictionaries. Meanwhile, the proposed penalty term within discriminative amplitudephase dictionary learning algorithm can profitably restrain the cross representation between signals and the proposed algorithm has a significant performance in signal separation.