In non-linear dynamic batch processes, the measured variables have different serial correlations, and the cross correlation among the variables could be reflected at different sampling moments, however, traditional detection methods do not consider the correlation among the variables, the relationships among all variables are usually regarded as independent or correlative for feature extraction, and the features of fault information are not fully extracted, so the monitoring effect is bad. Therefore, a batch process fault detection algorithm based on the kernel dynamic latent variable-dynamically weighted support vector data description (KDLV-DWSVDD) of variable blocks is proposed. Firstly, the variables are divided into related and independent variable sub-blocks through obtaining mutual information (MI) values among the variables. Then, KDLV algorithm is used to divide the related variable sub-block into a dynamic part and a static part, the vector auto-regressive model is established to monitor the dynamic part and the neighborhood preserving embedding (NPE) algorithm is used to monitor the static part. In the independent variable sub-block, DWSVDD algorithm can be used to extract the dynamic information of independent variables. Finally, the monitoring statistics are established for fault detection through KDLV-DWSVDD algorithm. The average fault detection rate of the proposed algorithm in the penicillin fermentation simulation process reaches 90. 38% , which is nearly improved by 15% compared with that of the comparison algorithms. The actual semiconductor industry process also proves the feasibility and superiority of the proposed algorithm for the fault detection of batch processes.