In order to improve the output accuracy of fiber optic gyroscope, the BP neural network model optimized by the beetle antennae search algorithm (BAS) was used as the base learner, and the Bagging parallel integrated learning algorithm was used to establish a BAS-BP-Bagging temperature compensation model, and a temperature compensation experiment was conducted for a certain model of fiber optic gyroscope. The experimental results show that under the temperature change environment from -40 ℃ to +60 ℃, the temperature drift of the fiber optic gyroscope after compensation is reduced by nearly 80% compared with that before compensation, 55% compared with the polynomial compensation algorithm, and about 30% compared with the BP neural network compensation algorithm. And the model shows superior generalization performance in the compensation of fresh samples.