低成本捷联惯导不对称动态误差的神经网络补偿

Application of Neural Network to Compensate Asymmetry Dynamic Errors in Low-cost SINS

针对低成本捷联惯导系统(SINS)中陀螺动态误差的不对称性在角振动条件下造成姿态漂移的问题,设计了多层前向神经网络的补偿模型。在标定模型参数时,为降低对外部参考信号测量精度的要求,提出用姿态解算的最终误差作为网络优化目标的训练方法。由于最终的姿态误差不是网络的期望输出,无法采用有导师的训练方法,为此采用了微粒群优化算法。仿真实验结果表明:补偿后的陀螺动态误差的不对称度减小了一个数量级。

In a low-cost strapdown inertial navigation system（SINS）, a multilayer feedforward neural network （NN） was designed to compensate the gyros asymmetry dynamic errors which caused attitude drift in rate oscillation. To reduce the accuracy demand of the reference signals in calibrating the NN model, the terminal attitude errors were computed as the network performance function for NN training. Unlike the supervised training, the terminal attitude errors were not the network target outputs. Therefore, the particle swarm optimization algorithm was applied to train the network. Simulation experiment results demonstrate that gyros asymmetry dynamic errors were reduced to about ten percent of those without the NN compensation.