Abstract: Transmitted in-seam wave seismic exploration is an effective method for detecting geological structures and hazardous bodies in working faces, but it suffers from problems such as shallow exploration depth, low resolution, and susceptibility to terrain and environmental noise. To address these issues, deep learning technology was introduced into transmitted in-seam wave seismic exploration to predict the location of strike-slip faults in working faces. A geological model of the strike-slip fault in the working face was established, and the elastic wave finite difference algorithm was used to perform forward modeling of in-seam waves to generate a simulation dataset. A Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) model was constructed. CNN was used to extract local features of the in-seam wave data, and LSTM was used to capture the temporal dependencies in the in-seam wave data, thereby achieving collaborative analysis of spatiotemporal features. The CNN-LSTM model was trained using the in-seam wave simulation dataset. The predicted root mean square error was 4.393 4 m, the mean absolute error was 2.987 5 m, and the coefficient of determination was 0.988 3, verifying the model’s high prediction accuracy and good generalization capability. The CNN-LSTM model was then fine-tuned using transfer learning and validated using transmitted in-seam wave exploration data from the 506 working face of a mine in Inner Mongolia. The results showed that the predicted fault location and strike were consistent with the actual position revealed by mining, and the prediction performance was better than that of the in-seam wave energy attenuation imaging and radio borehole penetration detection technologies.