Unlike continuous processes, a batch process contains many batch runs. Considering the repetitive nature of batch processes, an iterative learning strategy is proposed to estimate the state in batch processes, where the state prediction is updated twice rather than once in conventional state estimation methods: within a batch run, the measurements are employed to update the state prediction to obtain the state estimate; along the batch dimension, the estimation performance of pervious batch runs is used as a learning reference for the state estimate according to the repetitive nature. As a result, the current batch run is related with previous batch runs during the state estimation, and the information of the whole batch process is incorporated. Considering that the batch process is characterized by nonlinearity and non-Gaussianity, the particle filtering method is employed as the key algorithm for the state estimation. The effectiveness and practicability of the proposed method is indicated by its application in a beer fermentation process.