To address the issue of nonlinear impairments in optical fiber transmission for probability shaping （PS） system, a novel optical fiber nonlinearity compensation scheme integrating triplet and complex-valued convolutional neural networks（T-CCNN） is proposed. Initially, symbols processed through linear impairment correction are transformed into two-dimensional triplets. These triplets are then fed into the T-CCNN for fiber nonlinearity compensation, while preserving the correlation of phase information between the real and imaginary parts. The simulation results show that at the optimal transmit power, when transmitting over 800 km with a triplet side length of 13, the gain in the Q-factor of this scheme is 1.39 dB. Moreover, when reaching the hard decision forward error correction threshold, this scheme can increase the transmission distance by approximately 180 km compared to the DBP-1StPs scheme.