Aiming at defects of different denoising methods, a minimum description length denoising method based on the improved normalized maximum likelihood (INML) was proposed. This method encodes the model class and adding to the original code length and adaptively defines the threshold in signal denoising. Numerical simulation signals and experimental signals of rolling element bearings were used to test and compare the performances of the proposed method with VisuShrink, BayesShrink and RNML etc. The results showed that the INML-based noise cancellation method had more effective denoising performance and higher SNR, not only in eliminating random noise, but in preserving interested components of the signals. INML method enriched the theory of MDL denoising and enhanced its performance.