Strike-slip fault identification is vital to the exploration and development of ultra-deep fault-controlled fractured-vuggy reservoirs, but the seismic response of strike-flip fault is difficult to accurately identify on sections. Aiming at the problem that it is difficult to identify the strike-slip fault in Tarim Basin, we propose a deep-learning fault detection method based on U-net convolutional neural networks for seismic data. Considering the differences between synthetic data labels and actual fault fractures in the field seismic data, and the concept that deep-learning is data-driven method, we use the forward modeling to make training data and labels that conform to the seismic reflection wave characteristics of ultra deep strike-slip faults. We achieve precise descriptions of main and secondary low order faults without changing the network structure. Through testing field data, the network model accurately predicts the main strike-slip fault and secondary branch structures, the approach verifies the accuracy and noise resistance of convolutional neural networks for fault detection with seismic data, and reflects the great potential of deep-learning methods in the field of subtle description of oil and gas reservoirs.