Abstract: The environment for seismic exploration in underground coal mines is complex, and field data acquisition often leads to bad traces or missing data issues. To improve data quality, seismic trace interpolation is a key step in data processing. By adopting the dilated convolutional neural network (DLCNN) method based on the compressed sensing (CS) framework, and combining with the sparsity expression of CS and the nonlinearity of DLCNN method, deep learning interpolation techniques are applied in seismic data processing in coal mine working faces. This technology mainly consists of two parts: introducing the DLCNN framework to train a large number of image patches, iteratively optimizing input/output noise residuals to obtain DLCNN multi-scale filters, namely self-adaptive sparse dictionary elements; integrating the DLCNN part into the CS framework, using the DLCNN multi-scale convolution kernel as a learnable dictionary, and ultimately reconstructing missing seismic data using the CS framework to output interpolated results. The effectiveness of this method on theoretical missing seismic data was experimentally tested, the effectiveness of DLCNN nested CS technology was verified compared with the results of conventional linear interpolation and traditional CS interpolation reconstruction techniques. Applying this technology to the actual transmission data interpolation of 45205 working face in Sandaogou Coal Mine, and comparing the seismic data before and after interpolation, the strong signal in-phase axis continuity of the data was well maintained, this technology provided a new idea for seismic data interpolation processing in underground coal mines.