Different degrees of noise pollution in power system will affect the accuracy of harmonic detection. Traditional wavelet threshold de-noising methods have many defects in practical engineering applications, such as the hard threshold function is discontinuous at the threshold, the soft threshold function has fixed deviations, and the quantization rules lack of scale adaptability, etc. In view of this, we improve the wavelet threshold de-noising from two aspects of the threshold function and the threshold quantization rule. In this paper, the improved soft and hard threshold trade-off function combined with the layered threshold selection rule is used to de-noise the analog harmonic power signal. In the MATLAB environment, the simulated noise-affected power harmonic signal is de-noised using traditional and improved threshold methods. The quantitative indicators analysis is conducted from three de-noising effect. The results show that the improved threshold de-noising method improves the output signal-to-noise ratio of the reconstructed signal, which reduces the root mean square error and smoothness, and can flexibly and effectively de-noise without sacrificing computational complexity.