Aiming at the problem that the existing stage division strategy does not consider the dynamic and multi-stage characteristics of batch processes at the same time, resulting in poor process detection effect, an improved multi-phase kernel principal component analysis based on combined value function (CVF-MKPCA) algorithm is proposed. Firstly, the three-dimensional data of the batch process are expanded in the corresponding direction, and the dynamic characteristics between the data of the batch process are extracted by constructing the expansion matrix. Secondly, an improved combined value function is constructed to evaluate the structural similarity between different time series information; then, according to the evaluation requirements of dynamic structural similarity, the bottom-up search method is used for stage division, and MKPCA method is used for stage modeling. Finally, a new combine statistic is constructed to detect faults in each stage. In the simulation process of penicillin fermentation, the false alarm rate of the proposed algorithm is 3. 40% when the control limit is 95%, and 7. 98% when the control limit is 99%, compared with the comparison method, the false alarm rate is reduced by 2. 12% and 1. 26% respectively, which proves that the proposed method has better fault detection performance.