Position Correction Model Based on Gated Hybrid RNN for AUV Navigation

Abstract

Underwater autonomous navigation technology is a key link for AUV to complete various underwater missions successfully. In this paper, we have proposed a position correction model based on hybrid gated RNN, which does not need to establish a motion model like typical navigation algorithms, to avoid modeling errors in the navigation process. According to the low update rate of GPS and the slip data caused by the coverage of GPS antenna during the process of floating and diving, it does not meet the actual requirements of navigation and control. We proposed to use improved EKF and iSAM as the output of the training dataset to filter out GPS slips, then the position error is corrected at the same time. First, the normalized input sequence is fed to the two-layer hybrid gated RNN with BiLSTM output mode as sequence and LSTM output mode as last, and then the processed data is fed to the two fully connected layers with a dropout layer added in the middle. After that, the output of the position correction model can be obtained through the regression layer. To confirm the validity of the proposed model, we conducted sea trials to compare the proposed model with EKF and iSAM. The experimental results prove that the proposed position correction model performs better than EKF and iSAM in all aspects, and the correction model trained with iSAM is more effective than that trained with EKF.