The method of extracting Brillouin spectrum characteristics by using the general regression neural network is proposed based on the relationship between the optical fiber strain and the Brillouin spectrum frequency shift. The Brillouin spectrum frequency shift and gain are taken as the input vector and target vector of general regression neural network, respectively. Then the general regression neural network is trained and simulated. The more accurate Brillouin spectrum frequency shift can be calculated with the obtained weight and threshold. Experimental results and theoretical analysis show that the Brillouin spectral feature and optical fiber strain obtained by using the general regression neural network are more accurate compared with the nonlinear least square method、back propagation neural network and radial basis function network, and the corresponding optical fiber strain error is the least (within 1%).