Traditional genetic algorithms have convergence speed and premature easily in solving specific optimization problems, in case of system parameters identification, this paper presents an improved genetic algorithm. By selecting reasonable replication strategy, and improving the fitness function calculation, overcomes the prematurity, ensures the diversity of the population, avoid slow convergence which is resulted by the close fitness value in the later time. Solving the parameters of typical second-order system transfer function by the algorithm, in the case of larger SNR, obtained the estimation to be almost no bias. Experimental results showed the effectiveness of the method.