Aiming at the shortcomings of the present classification algorithms on density-based imbalanced data which are selected sequentially, an online sequential extreme learning machine based on granular division is proposed. In the offline stage, majority class samples are divided by using granularity according to the data distribution property, and the centre of granule is introduced for replacing the samples in this granule. In the online stage, the boundary majority samples are divided again by using granularity according to the new updated distribution property, and then are replaced by the new centre of granule to update network weight dynamically. Furthermore, a theoretical proof is given to testify the proposed algorithm had upper bound of information loss. The experimental results show that the proposed method can improve the total generalization performance and classification efficiency compared with some state-of-the-art algorithms.