Noisy Data Cleaning Method Based on Adaptive Wavelet Threshold

Due to the ability of multi-resolution analysis, the wavelet-based method has become an effective tool in signal denoising, and has achieved good results, but shortcomings still remain. In view of the problem that the accuracy of traditional thresholds is greatly affected by sampling parameters, while the hard threshold function introduces discontinuous and soft one presents bias, this paper puts forward an adaptive wavelet threshold denoising method, comprising a sliding energy window based threshold processing method (SEWBT) and an iterative filtering based noise variance estimation method. Based on the mathematics mechanics of the wavelet transform for white noise, SEWBT is proposed to replace classical threshold techniques and thus immunes from the discontinuous and bias. Afterward, an iterative filtering based method for the noise variance estimation combining SEWBT and Lilliefors test is presented. The method is validated with simulated and online measured partial discharge signals. The results show that the proposal is insensitive to sampling parameters or noise variance. While SEWBT outperforms traditional threshold functions as it can eliminate the white noise effectively, and maintain the waveform of pulses favorably.