Deep Learning Aided Magnetostatic Fields Based Real Time Pose Estimation of AUV for Homing Applicati

Autonomous underwater vehicles (AUVs) require a submerged docking station (DS) for recharging their batteries and data transfer purposes, in order to increase the vehicle endurance and updating its missions in the current subsea scenarios. Hence, a precise short-range guidance toward the DS is needed for AUVs for a reliable homing process in dynamic ocean environments. In this letter, a deep learning (DL) aided short-range guidance technique is proposed for reliable and precise homing of AUVs using an intelligent magnetic field guidance. A magnetic field is generated up to 20 m range from the dipole coil wound on the DS. The 64 bit microcontroller-based data acquisition system with a three-axis geomagnetic sensor embedded with DL hardware is developed as a prototype for estimating the AUV pose relative to the DS. A DL model is trained with the acquired magnetic field data with the coordinates of the AUV. The customized DL model is envisaged for predicting range and bearing angle simultaneously and has been developed with required hidden layers and activation functions. In this problem, the optimal DL model with two hidden layers and 64 neurons provides 94% accuracy in the prediction of AUV pose parameters i.e., range and bearing angle relative to DS. The DL model is embedded with the AUV hardware for real-time pose estimation of the AUV toward the DS. The proposed DL-based methodology is robust for AUV homing applications when compared to conventional approaches and machine learning techniques in dynamic environments.