Abstract: Aiming at the problems of low Wireless Fidelity (Wi-Fi) fingerprint location, cumbersome maintenance and high training cost, an indoor location method based on Channel State Information (CSI) and improved Multiple Kernel Extreme Learning Machine (FO-MKELM) was proposed. Firstly, the CSI amplitude difference and reconstructed phase information are fused in the preprocessing stage to reduce the influence of environmental noise. Secondly, in the off-line training stage, piecewise Quantum Particle Swarm Optimization (QPSO) is used to find the optimal parameters for the model to improve the positioning accuracy and generalization performance. Then, in order to suppress the impact of environment changes on positioning performance, the online incremental learning and forgetting mechanism is introduced, where some new data are added for incremental learning to continuously update the localization model, and the data validity period is set to forget old data to reduce adverse effects. Finally, in the online prediction stage, the model output is matched with the tag library data to obtain more accurate coordinates. The experimental results show that the static average errors of 0.39 m and 0.51 m are obtained in two different environments of open corridors and complex laboratories, respectively. Compared with other positioning methods, the proposed algorithm has improved positioning accuracy and long-term stability.