Aiming at the performance optimization problem of multi-pulse position modulation(MPPM) method applied to visible light communication, an MPPM transmission design scheme based on autoencoder model is proposed. The scheme uses fully connected network(DNN) and convolutional neural network(CNN) to build autoencoder models respectively. The encoder side realizes the generation of MPPM source symbols by adopting multi-stage training strategy and self-defined loss function regulation, the decoding end completes the functions of channel and MPPM signal feature learning through a network constructed by a fully connected layer or a one-dimensional convolution layer. The simulation results show that the bit error performance of MPPM transmission system based on self encoder can reach and slightly better than the performance of traditional maximum likelihood sequence detection.