It is difficult to obtain many priori samples during fault diagnosis of machine equipment. However, most of classification methods are not able to catch the latent information from the unlabeled fault samples. So an online semi-supervised learning algorithm is proposed. In the proposed method, online sequential extreme learning machine is extended from supervised learning to semi-supervised learning based on the tri-training algorithm. Three primitive classifiers are builded using the unprecise labeled samples and the classifiers are used to label the unlabeled samples temporarily. Then the samples with high confidence level are choosed to extend the real labeled samples. At the end, the classifiers are also updated via the new labeled samples. Experi- ments on the semi-supervised benchmark data sets show that the proposed algorithm could achieve a small difference on the classification accuracy and difference of the training time less than 1.2 times when the sum of training samples are same but the ratio of labeled samples to unlableled samples is different. The experiment test of fault diagnosis in a diesel is developed. The results show that the online semi-supervised learning algorithm could obtain an increase from 5% to 8% on the classification accuracy when the labeled samples are fewer but the unlabeled samples are added.