The invention discloses a data driving threshold value noise-reduction method for rotary machine vibration signals. According to the method, firstly, wavelet transformation is carried out on collected vibration acceleration signals, and the signals are decomposed into different frequency bands. Secondly, noise estimation is carried out on the signals at each frequency band, and data driving threshold values adapting to the signals are obtained. Thirdly, the signals are segmented through a sliding window technology. Finally, threshold value noise reduction is carried out on each segment of signals by utilizing the data driving threshold values, the signals are reconstructed, and time-domain signals after noise reduction are obtained. The data driving threshold values are from noise estimation of the signals, and the threshold values can be set in a self-adaptive mode according to the intensity of the noise. Compared with traditional threshold values, the data driving threshold values adapt to the signals, the threshold values are set more accurately, and weak fault signals are reserved while noise is reduced. The data driving threshold value noise-reduction method for the rotary machine vibration signals combines the advantages of wavelet transformation, the noise estimation algorithm, the sliding window technology and the 3sigma principle, can effectively extract failure characteristics, and achieves failure diagnosis of mechanical equipment.