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

Area of Science:

• Robotics and Autonomous Systems
• Navigation and Control
• Signal Processing

Background:

• Doppler velocity log (DVL) outages severely impact autonomous underwater vehicle (AUV) navigation, particularly in challenging environments like deep water or complex seabed terrain.
• Reliable navigation is critical for AUV mission success and safety.

Purpose of the Study:

• To develop a novel method for generating accurate pseudo-DVL velocity measurements during DVL outages.
• To improve the robustness and accuracy of loosely coupled inertial/Doppler navigation systems for AUVs.

Main Methods:

• An attention-guided temporal convolutional network (TCN) was employed to learn the mapping from inertial measurement unit (IMU) and strapdown inertial navigation system (SINS) data to DVL velocity.
• Causal dilated convolutions were utilized to capture temporal dependencies without relying on future information.
• An attention module was incorporated to prioritize informative motion segments for velocity prediction.

Main Results:

• Simulations across 16 AUV trajectories demonstrated that the proposed method outperforms baseline models like MLP, TCN, and GRU-Attention in reducing velocity and trajectory errors during continuous DVL outages.
• In a closed-loop turning scenario, the proposed method achieved a 14.69% reduction in eastward and a 14.74% reduction in northward velocity root mean square error (RMSE) compared to GRU-Attention.

Conclusions:

• The attention-guided temporal convolutional method effectively generates pseudo-DVL velocity measurements, significantly enhancing AUV navigation during sensor outages.
• This approach offers a promising solution for maintaining high navigation accuracy in AUVs operating in environments prone to DVL signal loss.