Sparse representation classification (SRC) has gradually received attention due to its powerful feature representation ability. However, the discriminative ability of traditional SRC methods is highly susceptible to the time-shift characteristic of vibration signals. To overcome the challenge, a multidomain kernel dictionary learning-based sparse classification (MDKDL-SC) method is proposed. First, a novel kernel discriminative dictionary (KDDL) is developed, in which a Gaussian edit distance with a real penalty kernel (ERP) is designed to tackle the time-shift property of the data. Second, a dictionary-based adjustable weighted voting strategy is developed in the recognition stage to leverage the representations learned from multiple domains. The weights of each domain are determined by a cross-validation method, which promotes the recognition performance by voting the weighted prediction label vectors. The performance of MDKDL-SC is validated by two datasets. Experimental results demonstrate that the MDKDL-SC method achieves the recognition rates of 99.75% and 99.52% in the two cases, respectively. Furthermore, the proposed method is compared with some cutting-edge methods, which further confirms the superiority of the MDKDL-SC method in machinery fault diagnosis.