Magnetic resonance sounding (MRS) signals are often severely distorted by noise, primarily the power-line harmonics which originates from anthropogenic interferences. Noise cancellation is a particularly important issue for the application of the MRS method. In order to reduce the harmonic peaks as well as the noise floor between the 50 Hz harmonics effectively, a method based on the wavelet modulus maximum is proposed. The method is based on marking the local modulus maximum points in the wavelet transform coefficients and removed the modulus maximum contributed by noise for MRS signal reconstruction. The optimum wavelet basis for noise cancelling is determined through simulations with synthetic signals. The accuracy and stability of the proposed method are investigated using synthetic signals embedded in both artificial and field-recorded noise. In addition, the results of the test on field data show the proposed method can effectively improve the fitting accuracy of the signal parameters and obtain stable inversion results.