A Data-Driven Method for INS/DVL Alignment

TL;DR

This work addresses the challenge of aligning inertial and Doppler velocity log sensors for long-range AUV navigation. It introduces AlignNet, a data-driven, 1D-CNN-based framework that estimates the rotation between the IMU and DVL frames from synchronized velocity measurements, replacing the traditional SVD-based method. Using a large simulated dataset of lawn-mower trajectories with realistic sensor noise, AlignNet achieves substantially faster convergence and lower RMSE than the baseline, demonstrating robustness to diverse sensor conditions. The results indicate that a learning-based alignment approach can significantly reduce pre-mission preparation time and improve navigation reliability in underwater environments, with plans for real-world sea tests.

Abstract

Autonomous underwater vehicles (AUVs) are sophisticated robotic platforms crucial for a wide range of applications. The accuracy of AUV navigation systems is critical to their success. Inertial sensors and Doppler velocity logs (DVL) fusion is a promising solution for long-range underwater navigation. However, the effectiveness of this fusion depends heavily on an accurate alignment between the inertial sensors and the DVL. While current alignment methods show promise, there remains significant room for improvement in terms of accuracy, convergence time, and alignment trajectory efficiency. In this research we propose an end-to-end deep learning framework for the alignment process. By leveraging deep-learning capabilities, such as noise reduction and capture of nonlinearities in the data, we show using simulative data, that our proposed approach enhances both alignment accuracy and reduces convergence time beyond current model-based methods.

Figures (4)

• Figure 1: AlignNet architecture showing the processing pipeline from input velocity measurements to estimated Euler angles. The network consists of three main components: input preprocessing, 1D CNN blocks with increasing feature dimensions (64, 128, 256), a global average pooling layer, and fully connected layers that output the final alignment angles.
• Figure 2: Overview of the AlignNet training pipeline. The ground truth velocity ($v^b_{GT}$) is processed through a noising pipeline to generate simulated DVL and INS velocities, which serve as input to AlignNet. The network estimates the alignment angles ($\hat{\theta}$), which are compared against the ground truth alignment angles ($\theta_{GT}$) using MSE loss to train the network.
• Figure 3: Simulation pipeline for generating DVL and INS velocity measurements. The upper path shows the DVL velocity generation process, where ground truth velocity is transformed and corrupted with noise and bias. The lower path shows the INS velocity computation through integration of noisy IMU measurements.
• Figure 4: Comparison of alignment RMSE performance between AlignNet and the baseline velocity-based SVD alignment method over time on simulated lawn mower trajectory (230 seconds). AlignNet reaches comparable accuracy within 25 seconds that the baseline method requires 100 seconds to achieve.