Attention-guided temporal convolutional pseudo-velocity generation for underwater inertial/Doppler navigation

Doppler velocity log (DVL) outages can rapidly degrade loosely coupled inertial/Doppler navigation for autonomous underwater vehicles, especially in deep water or over complex seabed terrain. This study presents an attention-guided temporal convolutional method for generating pseudo-DVL velocity measurements during such outages. The model uses recent inertial measurement unit outputs and strapdown inertial navigation system (SINS)-derived attitude, position and velocity as sequential inputs, and learns the mapping to horizontal DVL velocity measurements from DVL-available periods. During DVL-unavailable intervals, the trained model predicts pseudo-velocity measurements at the original DVL update instants, which are then used in an extended Kalman filter measurement update. Causal dilated convolutions are used to extract temporal dependencies without future information, while the attention module weights motion segments that are more informative for velocity prediction. Simulations using 16 autonomous underwater vehicle trajectories show that the proposed method reduces velocity and trajectory errors compared with multilayer perceptron (MLP), temporal convolutional network (TCN) and gated recurrent unit with attention (GRU-Attention) baselines under continuous DVL outage. In a closed-loop turning task, the eastward and northward velocity root mean square error (RMSE) values are reduced by 14.69% and 14.74%, respectively, compared with GRU-Attention.