The denoising of mechanical system is always an indispensable process in sensor signal analysis. It directly affects the result of subsequent tool state monitoring and identification. Therefore, a denoising framework is proposed to solve this problem. Bayesian nonparametric estimation instead of the Gaussian fitting distribution of CycleGAN can ensure the quality of denoising data to the greatest extent. The experiment of milling 42CrMo steel was carried out, and the proposed method was verified. Compared with the wavelet packet threshold, the signal-to-noise ratio (SNR) obtained by the propose model is increased by 4.71 dB on average, and RMSE ranges from 0.0210 to 0.0642. UKF-CycleGAN model has better denoising effect than other methods. The model proposed in this paper improves the accuracy of tool wear identification. At the same time, the process of selecting the parameters for denoising model by manual experience can be reduced. This provides the possibility for online denoising of sensor signals in milling process, which has certain guiding significance for tool state monitoring in machinery industry.