Abstract： In this paper, with the help of the rules of block matrix multiplication, an online semi-supervised extreme learning machine(OSSELM) was proposed according to semi-supervised extreme learning machine (SS-ELM)  based on manifold regularization.By the analysis of the manifoldregularization term of the objective function of SSELM, a kind of approximation algorithm of OSS-ELM named OSS-ELM(buffer) was proposed to avoid running out of memory in the process of online learning.The linear relationship between the sample number and the cumulative running time of the OSS-ELM(buffer) was revealed in the experiments using Abalone and the relative deviation of the generalization ability of the OSSELM and the SS-ELM is less than 1% in 9 public data sets, which show that the OSSELM(buffer) not only solves the problem of limited memory, but also improves the speed of online learning while keeping the generalization ability of SSELM. This proves that the OSSELM(buffer)can be effectively applied to online semisupervised learning.