Variational Autoencoder for the Prediction of Oil Contamination Temporal Evolution in Water Environments.

Area of Science:

• Robotics and Environmental Science
• Artificial Intelligence and Machine Learning

Background:

• Water quality monitoring of large water bodies using robotic vehicles is crucial but challenging due to dynamic and unknown environments.
• Current methods often rely on adaptive path planning and machine learning, but struggle with environmental unpredictability.

Purpose of the Study:

• To develop a dynamic contamination model for water quality assessment using partial observations from autonomous surface vehicles.
• To address the challenge of characterizing water quality in highly dynamic and uncertain aquatic environments.

Main Methods:

• A variational autoencoder (VAE) was developed and trained in a model-free manner.
• An oil spillage simulator was used for heuristic-based world building and training data generation.
• The VAE was tested with homogeneous fleets of autonomous surface vehicles equipped with various sensors in diverse simulated environments.

Main Results:

• The proposed VAE demonstrated accurate future contamination distribution predictions.
• Mean squared error ranged from 3% to 9% compared to a static baseline approach.
• The VAE exhibited high robustness against unseen scenarios, indicating low overfit.

Conclusions:

• The developed variational autoencoder provides an effective method for dynamic water quality modeling from partial observations.
• The approach enhances the prediction of contamination distribution, outperforming static methods.
• The model shows significant robustness and adaptability in complex environmental monitoring tasks.