The invention discloses a neural network construction system and method based on a variable-length genetic algorithm, wherein the construction system includes a sequentially connected training set generation module, a network initialization module, a network training module, a network update module, a network optimization module, Network selection module, iteration number update module and network generation module. The general idea of the construction method of this scheme is to generate variable-length chromosomes in the initialization step, and randomly add BN components to specific genes of each chromosome. Through the training on the training set, the neural network structure representation corresponding to each chromosome is obtained through training, and the offspring chromosomes are selected. Subsequently, in the crossover step, the length of the crossed chromosomes is not fixed, and a growth and shrinking strategy is used to generate daughter chromosomes. Then use the traditional mutation and environment selection operations to complete the selection of offspring chromosomes, repeat the above, and decode the obtained best chromosomes into the corresponding neural network architecture.