[Paper Review] DANAE: a denoising autoencoder for underwater attitude estimation

Paolo Russo, Fabiana Di Ciaccio+1 more|arXiv (Cornell University)|Nov 13, 2020

Underwater Vehicles and Communication Systems29 references4 citations

TL;DR

This paper proposes DANAE, a deep denoising autoencoder that improves underwater attitude estimation by reducing noise in Kalman filter outputs from IMU/AHRS data. By learning to reconstruct clean Euler angles from noisy inputs, DANAE reduces RMSE by 63% on OxIOD and 55% on UCS, demonstrating robust, filter-agnostic noise compensation without hyperparameter tuning.

ABSTRACT

One of the main issues for underwater robots navigation is their accurate positioning, which heavily depends on the orientation estimation phase. The systems employed to this scope are affected by different noise typologies, mainly related to the sensors and to the irregular noise of the underwater environment. Filtering algorithms can reduce their effect if opportunely configured, but this process usually requires fine techniques and time. In this paper we propose DANAE, a deep Denoising AutoeNcoder for Attitude Estimation which works on Kalman filter IMU/AHRS data integration with the aim of reducing any kind of noise, independently of its nature. This deep learning-based architecture showed to be robust and reliable, significantly improving the Kalman filter results. Further tests could make this method suitable for real-time applications on navigation tasks.

Motivation & Objective

• To address the challenge of noisy attitude estimation in underwater robotics, where GPS is unavailable and sensor noise degrades navigation accuracy.
• To develop a noise-robust, filter-agnostic deep learning model that compensates for both random and bias errors in IMU/AHRS data without requiring prior knowledge of noise characteristics.
• To improve the performance of Kalman filter-based attitude estimation by leveraging denoising autoencoders to reconstruct cleaner Euler angle signals.
• To evaluate the method on real-world underwater datasets and demonstrate its potential for real-time deployment in autonomous underwater navigation.

Proposed method

• DANAE uses a deep 1D convolutional encoder-decoder architecture with dilated convolutions to learn a compressed latent representation of noisy IMU/AHRS input sequences.
• The model employs a denoising autoencoder objective, training on corrupted input sequences to reconstruct clean Euler angles (roll, pitch, yaw) from the original signal.
• The decoder uses transposed-dilated convolutions to upsample the latent representation back to the original sequence length, with skip connections combining corresponding encoder and decoder layer outputs.
• The architecture is trained end-to-end using the Adam optimizer with a fixed learning rate of 0.002 and a batch size of 16 over 100 epochs on UCS and 1 epoch per OxIOD subset.
• The method is filter-agnostic, meaning it can be applied to outputs from any Kalman filter variant (e.g., EKF, UKF) without modification.
• The model is evaluated using standard metrics: mean deviation, maximum deviation, and RMSE, compared to ground truth.

Experimental results

Research questions

• RQ1 Can a deep denoising autoencoder effectively reduce noise in Kalman filter outputs for underwater attitude estimation without prior knowledge of noise sources?
• RQ2 How does the performance of DANAE compare to the Kalman filter baseline in terms of RMSE, mean deviation, and maximum deviation across diverse underwater datasets?
• RQ3 To what extent is DANAE robust to varying noise characteristics, including sensor bias and environmental interference, without requiring hyperparameter tuning?
• RQ4 Can DANAE be applied effectively to nonlinear Kalman filters (e.g., EKF, UKF) and scaled for real-time underwater navigation?

Key findings

• On the OxIO dataset, DANAE reduced the RMSE for roll, pitch, and yaw by 63%, 68%, and 45% respectively compared to the Kalman filter baseline.
• On the UCS dataset, DANAE reduced the RMSE for roll and pitch by 55% and 54% respectively, with mean deviation reduced by 57% and 55%.
• The maximum deviation was reduced by 53% for roll and 49% for pitch on the OxIO dataset, indicating improved robustness to outliers.
• The model showed consistent performance across different noise profiles and did not require re-tuning for different datasets, demonstrating strong generalization and robustness.
• Visual comparisons in Figure 1 and Figure 2 show that DANAE's output closely follows the ground truth, significantly reducing signal elongation and jitter compared to the Kalman filter.
• The yaw angle remained unstable in both KF and DANAE due to magnetometer nonlinearity and environmental interference, but DANAE still reduced its RMSE by 45% on OxIO and 54% on UCS.