Extreme Learning Machine (ELM) is a novel single-hidden-layer feed forward neural network with fast learning speed and better generalization performance compared with the traditional gradient-based learning algorithms. However, ELM has two issues: the hidden node number of ELM needs to be predefined and the random determination of the input weights and hidden biases lead to ill-condition problem. In this paper, a two-stage evolutionary extreme learning machine (TSE-ELM) algorithm was proposed to overcome the drawbacks of original ELM, which used an improved artificial bee colony (ABC) algorithm to optimize the input weights and hidden biases. The proposed TSE-ELM algorithm was applied on the UCI benchmark datasets and rolling bearing fault diagnosis. The numerical experimental results demonstrated that TSE-ELM had an improved generalization performance than traditional ELM and other evolutionary ELMs.