Abstract: In order to obtain an accurate constraint on the effective mass splitting of protons and neutrons in nuclear matter, a dual-channel-input convolutional neural network (CNN) is proposed. The data sets for learning, the longitudinal and transverse momentum distribution of protons and neutrons , are generated from improved quantum molecular dynamics (ImQMD) model with two kinds of Skyrme parameters with opposite nucleon effective mass splitting, i.e., SkM * and SLy4. The results show the accuracies beyond 99.5% in 50 MeV/u for 3 systems, ⁴⁸Ca+²⁰⁸Pb, ⁴⁸Ca+¹²⁴Sn and ¹²⁴Sn+¹²⁴Sn. When the beam energy is 270MeV/u, the accuracies of the 3 systems are still higher than 93%. Among them, the accuracy of ⁴⁸Ca+²⁰⁸Pb is the best, which is 98.6%. Finally, the importance maps of of 3 systems at 50MeV/u are provided. It is indicated that the two-dimensional energy spectrum of nucleon in low region is more sensitive to the effective mass splitting.