This paper proposes an attention-based Multi-level wavelet Convolutional Neural Network to suppress free surface multiples in marine seismic data. Wavelet transform is used to compress the feature size of image data to avoid the loss of information caused by traditional down-sampling. Besides, it also introduces an attention mechanism to expand its receptive field and improve the fidelity of training. The algorithm proposed in this paper is compared with DnCNN network and U-Net network to test the simulated data and actual data under different observation modes. The experimental results show that the attention Mechanism in MWCNN can better separate the primary wave and the free surface multiple, and the protection of the effective signal are better than the other two networks. It has strong generalization ability and suppression efficiency.