An implicit nonlinear error model is established for the measurement errors of three-axis accelerometers

and an auto reverse tuning extreme learning machine（RT-ELM） is proposed to train the error model. The experimental results show that the error after compensation is basically controlled within ± 0. 07 m/s~2

the root-mean-square error is less than 0. 004 m/s~2

and the error is reduced by more than one hundred times before the compensation

which is higher than the fixed ELM about 7 times. Compensation effect are basically the same

selected training set and test set randomly

it proves that the extreme learning algorithm has good generalization ability and robustness. The training time for thousands of sample points is only about 0. 06 s

and its speed is thousands of times higher than that of traditional back propagation（BP） neural networks. It is suitable for fields such as error compensation and control systems that require high real-time performance.