Abstract: Noise components exist in the actual signals obtained by data feature acquisition in the running process of CNC machine tools, and the actual running status of machine tools cannot be judged according to the signal features. In order to improve the ability of signal denoising extracting characteristic parameters, a spectral wavelet threshold denoising (SWTD) algorithm was constructed according to the spectral wavelet transform theory and practical application process, which could analyze one-dimensional digital signals. Then, the reliability of spectral wavelet threshold denoising was verified by the denoising method of simulation signal and machine tool spindle vibration signal. The research results show that the useful frequencies contained in the vibration signals of machine tool spindle are basically in the low frequency region within the range of 400 Hz. After the denoising is completed, the SWTD denoising signal forms a stable amplitude, which is close to the initial signal, and the high frequency band of more than 400 Hz in the envelope spectrum is eliminated, forming the main component consisting of the spindle rotation frequency and the hobbing frequency. Obvious characteristics of the spindle rotation frequency are observed, indicating that the proposed method has better performance than the wavelet threshold denoising results.