Supersonic civil aircraft represent an important development direction for the new generation of civil aircraft. Their unique sonic boom phenomenon is the most crucial factor restricting supersonic flight over land. In engineering, the sonic boom prediction method combining computational fluid dynamics (CFD) and acoustic propagation theory is mainly used. The near-field overpressure distribution is a key input parameter for calculating the ground sonic boom waveform, and its accuracy directly affects the far-field sonic boom prediction results. To alleviate the contradiction between the cost and accuracy of sonic boom prediction, this paper adopts a multi-fidelity transfer learning strategy based on multi-channel fusion (MF-TLNN) to establish the mapping relationship be-tween the aerodynamic shape and the near-field overpressure distribution. This strategy effectively utilizes low-fidelity and high-fidelity data by combining the concepts of multi-fidelity data fusion and transfer learning, improving pre-diction accuracy and reducing overfitting and negative transfer problems. The research results show that compared with the single-fidelity deep neural net-work, the MF-TLNN model can achieve low-cost and high-accuracy near-field overpressure distribution prediction by jointly using low-fidelity data under the condition of limited high-fidelity data, providing a good foundation for sub-sequent far-field sonic boom prediction and guiding the aerodynamic shape optimization for low sonic boom.