The neighborhood preserving embedding(NPE) algorithm can preserve the data local structure, but ignore the data global structure for fault detection of complex dynamical batch process, which causes low detection rate. An improved NPE algorithm based on cross entropy global-local neighborhood preserving embedding(CEGLNPE) is proposed. Firstly, the idea of the global structure of cross entropy is introduced into NPE, and the objective function with global-local structure is obtained. Then the sliding window is used to further solve the dynamic problem of batch process and a global-local fault detection model is established to improve fault detection efficiency and accuracy, which can preserve the structures of both global and local data simultaneously. Finally, SPE and $ {T^{2