In complex exploration environments, the raw seismic data collected contains a large amount of random noise, which seriously affects the quality of seismic data and brings difficulties for subsequent geological interpretation. In response to this issue, a seismic signal denoising model combining deformable convolution and attention mechanism has been proposed to solve the above problems, namely DnDCNN (denoising deformable convolutional neural network). Firstly, the attention mechanism that integrates deformable convolutions was introduced into DnCNN, making the network more focused on the effective signal area and reducing the loss of detailed information. Secondly, the concatenated standard convolution was replaced with deformable convolution and standard convolution concatenated mode, which improves the network's ability to extract invariant features. Finally, batch normalization and residual learning strategies were integrated to achieve fast network convergence and signal-to-noise separation. The validation of simulated and actual earthquake data has shown that the network model can effectively suppress random noise, preserve more detailed information, and exhibit better signal-to-noise ratio in denoising weak signals at different noise levels.