The invention discloses a reproduction device for a rocket sled track spectrum and a method thereof. A deep bidirectional cyclic neural network is used to construct a machine learning model, a plurality of groups of white noise excitation signals are constructed as model output, the excitation signals are applied to a bracket-track system and acceleration signals of one point on the track are collected as model input, and the model is trained and a structural response-excitation model is obtained; track random irregularity simulation is carried out based on a power spectral density function, and spatial power spectral density is converted to time domain signals by inverse discrete Fourier transform; expected spectrum signals are converted to acceleration signals and are inputted to a learning model to obtain expected excitation signals, the excitation signals are applied to the bracket-track system, and response point acceleration signals are collected and compared with an expected waveform for realizing track spectrum reproduction. Through setting the parameters of the algorithm and the layer number of the network structure, features of the structure can be better acquired, and anexcitation applying scheme with high accuracy can be given.