Aiming at the problem of difficult feature extraction of rolling bearing vibration signals in the strong noise background, based on the basic theory of sparse representation, a multi-regularized sparse reconstruction noise reduction model using multi-factor evolutionary algorithm is proposed. Firstly, the solution of the multi-regularization model is divided into three more objective subtasks, the l_0-paradigm constrained optimization main task and the l_1 and l_p-paradigm regularization additional tasks, and the above tasks constitute three different objectives of the sparse reconstruction algorithm for multi-factor optimization; secondly, according to the priority of different regularization tasks in the evolutionary process, the golden segmentation search strategy is used to ensure that each community contains individuals with similar fitness, and the sparsity characteristics of the samples are guaranteed by the two-point crossover genetic operator; lastly, the thresholding iterative algorithm is applied to the local search process to accelerate the population convergence in the subtask. On this theoretical basis, the feasibility of this method is verified by simulation signals and actual bearing data, respectively. The experimental results show that the method can effectively extract the impact features under strong noise background and provide reliable a priori knowledge for further fault diagnosis.