K-Hybrid: A Kurtosis-Based Hybrid Thresholding Method for Mechanical Signal Denoising

This paper presents a kurtosis-based hybrid thresholding method, K-hybrid, for denoising mechanical fault signals. The threshold used in the hybrid thresholding method is determined based on kurtosis, which is an important indicator of the signal-to-noise ratio (SNR) of a signal. This together with its sensitivity to outliers and data-driven nature makes a kurtosis-based threshold particularly suitable for on-line detection of mechanical faults featuring impulsive signals. To better reflect the signal composition, the proposed hybrid thresholding rule divides the wavelet transformed input signals into four zones associated with different denoising actions. This alleviates the difficulties present in the simple keep-or-remove and shrink-or-remove approaches adopted by the hard- and soft-thresholding rules. The boundaries of the four zones are on-line adjusted in response to the kurtosis change of the signal. Our simulation results suggest that the mean squared error (MSE) is unable to distinguish the results in terms of the amount of falsely identified impulses. It is therefore inappropriate to use MSE alone for evaluating the denoising results of mechanical signals. As such, a combined criterion incorporating both MSE and false identification power $Pfalse$ is proposed. Our analysis has shown that the proposed K-hybrid approach outperforms the soft, hard, and BayesShrink thresholding methods in terms of the combined criterion. It also compares favorably to the MAP thresholding method for signals with low kurtosis or low SNR. The proposed approach has been successfully applied to noise reduction and fault feature extraction of bearing signals.