Abstract: Due to complex environmental and economic constraints, seismic records often contain missing traces, which could cause some challenges for subsequent seismic signal processing tasks. In recent years, owing to the outstanding feature extraction capabilities of Convolutional Neural Networks (CNNS), CNN-based methods have achieved a certain level of success in seismic trace reconstruction tasks. However, due to the limitation of the CNN's local receptive field, CNN-based methods focus on extracting local features and ignore global features. In addition, supervised methods for training networks require high quality data pairs, which is a challenging task for 3D seismic data. To solve the above problems, this paper proposes a reconstruction algorithm STC-Net based on self-supervised training combined with Swin Transformer and CNN. Through the extraction and fusion of global and local features, incomplete 3D seismic data could be reconstructed. Experimental results show that the proposed global-local feature extraction algorithm has better signal reconstruction capability compare with traditional methods and CNN-based methods.