A data-driven threshold for wavelet sliding window denoising in mechanical fault detection

Wavelet denoising is an effective approach to extract fault features from strong background noise. It has been widely used in mechanical fault detection and shown excellent performance. However, traditional thresholds are not suitable for nonstationary signal denoising because they set universal thresholds for different wavelet coefficients. Therefore, a data-driven threshold strategy is proposed in this paper. First, the signal is decomposed into different subbands by wavelet transformation. Then a data-driven threshold is derived by estimating the noise power spectral density in different subbands. Since the data-driven threshold is dependent on the noise estimation and adapted to data, it is more robust and accurate for denoising than traditional thresholds. Meanwhile, sliding window method is adopted to set a flexible local threshold. When this method was applied to simulation signal and an inner race fault diagnostic case of dedusting fan bearing, the proposed method has good result and provides valuable advantages over traditional methods in the fault detection of rotating machines.