The MR-DCA based rolling bearing fault diagnosis method is proposed for the problem that rolling bearing weak faults are difficult to identify. The input samples are pre-processed by using the maximun correlated kurtosis deconvolution and resonance-based sparse signal decomposition, which can effectively filter out the noise of original signal and feature the fault impact components. The obtained two-dimensional time-frequency diagrams of the fault components and the original signal are used as the training samples of the network, and after two feature learning modules, the input features are filtered using the attention mechanism, and the model computational efficiency and recognition accuracy can be effectively improved by weight reassignment. In order to verify the model performance, a rolling bearing weak fault dataset is used for fault diagnosis analysis, while ablation experiments are carried out to verify the effectiveness of each module of the diagnostic model. The results show that the proposed model can achieve good diagnostic performance in the diagnosis of rolling bearing weak faults.