Rolling bearings are one of the most important components in rotating machinery. During operation, the transient impact signals of their early failures are often masked by strong background noise, and traditional methods are difficult to effectively separate features. This paper proposes a multi-scale sparse maximum correlation kurtosis deconvolution method for extracting weak fault features of rolling bearings in strong noise environments. Firstly, using the Boltzmann Shannon interaction entropy as the optimization criterion for the variational modal decomposition (VMD) algorithm, the penalty factor αand mode number K of VMD are optimized through the GOOSE algorithm (GOOSE) to achieve adaptive decomposition of the signal, and the optimal mode components are selected based on the maximum kurtosis principle of the envelope spectrum. Secondly, in order to enhance the adaptability of signal pulse features, a new feature index KG was constructed, which integrates the sparsity measures of signal envelope spectrum kurtosis and pulse mode decomposition GM2tol, achieving multi-scale fusion of signal impulse components and global pulse distribution. With the feature index KGas the optimization objective, the filter length L, period T, and displacement M of the maximum corresponded kurtosis deconvolution (MCKD) parameter filter were optimized to enhance the features of the optimal component impulse components. Finally, the characteristic frequency of bearing faults is extracted through envelope spectrum analysis. The method was validated using simulated signals and experimental datasets, and the results showed that it clearly extracted fault frequencies and their multiples, with a fault frequency extraction error of 0.54%, achieving the discrimination of fault categories.