Abstract: ‍ ‍The drone RF signal identification technology is the key technology to realize automatic drone identification in next-generation air traffic control systems. Aiming to improve the drone radio frequency signals identification accuracy under low signal-to-noise ratios， this paper proposed a radio frequency signal identification model based on the multi-branch convolution structure and attention mechanism. Firstly， a lightweight backbone network composed of the multi-branch convolution structure was designed to extract the small time-varying features in the signal time-frequency spectrum. Secondly， the attention mechanism was used to enhance the expression of the signal region in the time-frequency spectrum， restrain the attention to the noise region， and further improve the signal identification accuracy. This paper carried out experiments on two open datasets to evaluate the performance of the model. The comparative experimental results show that the proposed model can effectively improve the identification accuracy of drone radio frequency signals under low signal-to-noise ratio conditions. The identification accuracy can reach 94.63% and 94.75% respectively in the low SNR range of -15 dB to -6 dB， and 97.50% and 97.35% respectively in the full signal-to-noise ratios range of -15 dB to 15 dB， which is significantly improved compared with the previous methods. The results of ablation experiments show that the multi-branch structure and attention mechanism can improve the performance by 3.74% and 1.56% respectively. The inference speed test experiment shows that the inference time of this model is only 1.61 ms， which can be used in the actual drone signals identification systems.