Multiwavelet denoising using SureBlock thresholding is proposed in order to effectively extract weak fault features which are immersed in noises cause by equipment and the surrounding environment. The effect of wavelet block threshold denoising mainly depends on the selection of shrinkage functions and thresholds. Multiwavelets have more than two different multiscaling functions and they possess excellent properties of the orthogonality, symmetry, compact support and high vanishing moments simultaneously. Based on the correlation between multiwavelet coefficients, this paper selects the optimal block length and threshold by using the minimum of Stein’s Unbiased Risk Estimate when estimating unknown fault features. The optimal block length and threshold are applied for the effective feature extraction and noise elimination at each decomposition level. The simulation signal validates the effectiveness of the proposed method, and the gearbox fault diagnosis in the rolling mills indicates that the proposed method can successfully extract two local scuffing faults caused by the surface welding at a high temperature on the pinion.