The simultaneous-source data acquisition technique allows seismic data to overlap with each other, and by exciting two or more sources simultaneously or delayed, it is possible to obtain several times more seismic data than conventional acquisition in the same time, which greatly improves the acquisition efficiency, but because the acquisition data are mixed with a large amount of confounding noise, it seriously affects the subsequent data processing and interpretation. This paper proposes a deblending method based on improved U-Net, which incorporates a dual channel attention mechanism into the original U-Net, focusing on the continuity of the reflection layer and waveform amplitude changes in seismic data, while enhancing the signal contrast in local areas and highlighting the reflection signals; The use of hybrid dilated convolution avoids partial information loss caused by pooling operations during down-sampling, ultimately achieving mixed data separation based on dual attention mechanism and hybrid dilated convolution U-Net (HDC AU-Net). The simulation data experiment results show that compared with the iterative sparse inversion method and the original U-Net method, the HDC AU-Net method has better removal effect on aliasing noise and higher separation signal-to-noise ratio. The actual data experiment further verified the reliability of the algorithm.