Abstract: To solve the problem of concept drift in complex industrial process and to improve the generalization performance of ensemble learning model, an ensemble modeling method of online dynamic selection for optimizing the diversity of the base learners was proposed, on the basis of ensuring the accuracy of the ensemble learning model. Online sequential extreme learning machine was used as the base learner, and the base learners were sorted in reverse order according to their classification accuracy on the sliding window. The other performance indexes of the basic learners on the sliding window were used as the feature attributes, and the approximate linear dependence condition was used to select accurate and diverse base learners for ensemble output, which improves the classification accuracy of the ensemble algorithm in dealing with the concept drift data stream. Finally, the rationality and effectiveness of the proposed algorithm were verified by using the synthetic data sets and real-world data sets.