In this article, we use a convolutional autoencoder neural network to reduce data dimensioning and rebuild soliton dynamics in a passively mode-locked fiber laser. Based on the particle characteristic in double solitons and triple solitons interactions, we found that there is a strict correspondence between the number of minimum compression parameters and the number of independent parameters of soliton interaction. This shows that our network effectively coarsens the high-dimensional data in nonlinear systems. Our work not only introduces new prospects for the laser self-optimization algorithm, but also brings new insights into the modeling of nonlinear systems and description of soliton interactions.