In the field of data test, environment vibration signals induced by rail transit and background vibration signals are inseparable. A new adaptive neuro-fuzzy inference system (ANFIS) method is put forward. Basic principle to banish background vibration is discussed, and the implementation steps of ANFIS are determined. A background vibration acceleration time history is superimposed on a ground vibration induced by a train to synthesize a testing field vibration record, which is used to remove background vibration by the proposed ANFIS method and other existing methods. The comparative analysis of results is carried out. The root mean square errors of time history curves calculated by these methods are 0.414 mm/s2 with the Fourier amplitude revising method, 0.363 mm/s2 with the auto power spectral density(PSD) method, 0.261mm/s2 with auto cross PSD method and 0.074 mm/s2 with the proposed ANFIS method, respectively. The error of ANFIS method is minimal. Also the weighted vibration level VLz values are 63.842 dB with vibration level revising method, 62.894 dB with Fourier amplitude revising method, 63.859 dB with auto PSD method, 63.802 dB with auto-cross PSD method and 63.805 dB with ANFIS method, respectively. The calculation value of ANFIS method is the closest to real value 63.815 dB. The results show that the time history, Fourier spectrum, power spectral density and vibration level obtained with the new ANFIS method are extremely close to true ones of traffic vibration, and the errors are relatively smaller than those of other existing approaches.