To solve the problem of on-line updating for the classification model in the case of fault feature samples added in batches, a dynamic sequence extreme learning machine(LSSELM) is proposed. Dynamic adaptability of LSSELM is improved by adding a new feature sample, and meanwhile, abandoning an old one with the highest similar degree of the same label to it iteratively. Then, the Sherman-Morrison formula is used to decrease the calculation complexity, and the output weights are solved recursively, and the online training of classification model is completed. Finally, applying LSSELM to online fault diagnosis of the analog circuit, the simulation results show that the LSSELM can get higher diagnosis accurate and better generalization than the online sequence extreme learning machine(OS-ELM).