Radar signal deinterleaving is to deinterleave the radar pulse description words obtained by electronic reconnaissance, which is the basis for radar recognition. At present, the effect of radar signal deinterleaving is mainly affected by factors such as parameter variation, measurement errors, high pulse density and high pulse loss. Radar signal deinterleaving methods can be divided into conventional methods and intelligent methods. Among the conventional deinterleaving methods, the research on pre-deinterleaving mainly focuses on the clustering algorithm itself, and the commonly used algorithms include K-Means, fuzzy C-means, data field, DBSCAN, graph theory and self-organizing neural network. The representative methods of main deinterleaving include histogram methods, spectral transform methods, correlation matching methods, plane transform methods, Kalman filter methods and multi-station time difference methods. The representative methods for the condition of known parameters is sample graph methods. The tools used in the intelligent deinterleaving method include supervised neural network, support vector machine, automata and Markov chain, etc. The relevant researches mainly focus on methods with supervised neural networks, and the representative methods include the deinterleaving methods using multiple neural network to deinterleaving by iteration, the deinterleaving methods based on image data semantic segmentation and the deinterleaving methods based on sequence data semantic segmentation. In the future, the conventional deinterleaving methods still has room for development in the direction of improving the clustering adaptability and the new main deinterleaving theory. The intelligent deinterleaving methods can be developed in the direction of the efficient deinterleaving neural network, few-shot deinterleaving model training and the deinterleaving under the condition of anti-intelligence. The recognition effect-oriented deinterleaving method is also worth exploring.