So far, there is no complete and unified theoretical guidance on the selection of BP neural network structure. This can only be chosen through experience. There are too many choices of network structure, training efficiency is not high or overconfiguration may occur, thus reducing network performance and fault tolerance. If the choice is too small, the network may not converge. The network structure directly affects the approximation ability and generalization of the network. Therefore, how to choose the appropriate network structure is an important problem for applications.