LSTM网络提高MEMS惯导定位精度的分析及验证

Analysis and verification on improving MEMS navigation accuracy based on LSTM network

为了提高GPS信号短期丢失状态下微机电(MEMS)惯导的导航定位精度,提出一种基于长短期记忆网络(LSTM)预测和传统GPS/MEMS组合导航系统相结合的高精度定位方法,对存在GPS信号状态下的LSTM模型进行训练来预测输出GPS信号丢失时的定位信息。针对单纯MEMS惯导推算误差发散快和反向传播神经网络(BPNN)无法处理时间序列数据的问题,采用LSTM来进一步抑制惯导累积误差,并使用自适应时刻估计方法来优化训练过程以提高模型性能。60 min时长的行驶测试数据集的验证结果表明:基于LSTM的MEMS惯导定位方法能够有效提高无GPS信号状态下的定位精度,相比于单纯MEMS惯导推算和BPNN的定位精度分别提高了94.62%和73.03%。

In order to improve the navigation positioning accuracy of MEMS in short-term absence of GPS signals,a high-precision positioning method based on the combination of long short-term memory (LSTM) prediction and traditional GPS/MEMS integrated navigation system is proposed.LSTM models are trained in the presence of GPS signals to predict positioning information when GPS signals are lost.To solve the problem that the error estimated by pure MEMS has fast divergence speed,and Back Propagation Neural Network (BPNN)are unable to process the time series data,LSTM is used to further suppress inertial cumulative errors,and an adaptive moment estimation is used to optimize the training process to improve the model performance.The verification results of the 60min driving test show that the LSTM-based MEMS positioning method can effectively improve the positioning accuracy in the absence of GPS signals,which is 94.62% and 73.03% higher than the positioning accuracies ofpure-MEMS and BPNN,respectively.