Abstract: Simultaneous source technology, which reduces seismic survey time and improves the quality of seismic data by firing more than one source with a narrow time interval, is compromised by the massive blended interference. Therefore, deblending algorithms have been developed to separate this interference. Recently, deep learning (DL) has been proved its great potential in suppressing the interference. The most popular DL method employs neural network as a filter to attenuate the blended noise in an iterative estimation and subtraction framework (IESF). However, there are still amplitude distortion and blended noise residual problems, especially when dealing with weak signal submerged in strong interference. To address these problems, we propose a hybrid WUDT-NAFnet, which contains two sub-networks. The first network is a wavelet based U-shape deblending transformer network (WUDTnet), incorporated into IESF as a robust regularization term to iteratively separate the blended interference. The second network is a nonlinear activate free network (NAFnet) designed to recover the event amplitude and further suppress the weak noise residual in IESF. With the hybrid network, the blended noise can be separated purposefully and accurately. Examples using synthetic and field seismic data demonstrate that the WUDT-NAFnet outperforms traditional curvelet transform (CT) based method and the deblending transformer (DT) model in terms of deblending. Additionally, for field applications, the data augmentation method of bicubic interpolation is applied to mitigate the feature difference between synthetic and field data. Consequently, the trained network exhibits strong signal preservation ability in numerical field example without requiring additional training.