Recurrent neural networks for hydrodynamic imaging using a 2D-sensitive artificial lateral line.

Area of Science:

• Robotics and Biomimicry
• Hydrodynamics and Sensor Technology
• Artificial Intelligence and Machine Learning

Background:

• The lateral line system in aquatic animals is crucial for sensing hydrodynamic stimuli.
• Developing artificial systems to mimic this biological capability is key for underwater navigation and object detection.
• Existing methods for artificial lateral line (ALL) source localization often use feed-forward neural networks.

Purpose of the Study:

• To develop and evaluate a 2D-sensitive artificial lateral line (ALL) system for object localization.
• To compare the performance of recurrent neural networks (LSTM) against feed-forward neural networks (OS-ELM) for hydrodynamic sensing.
• To assess the impact of noise on localization accuracy.

Main Methods:

• An artificial lateral line (ALL) with eight all-optical flow sensors was constructed.
• Hydrodynamic velocity profiles were measured in response to a moving 6 cm sphere.
• Object location was reconstructed using both Long Short-Term Memory (LSTM) and Online-Sequential Extreme Learning Machine (OS-ELM) neural networks.

Main Results:

• The LSTM achieved a significantly lower average localization error (0.72 cm) compared to the OS-ELM (4.27 cm) within a 62 cm x 9.5 cm area.
• The recurrent neural network (LSTM) demonstrated greater resilience to noise in the measurements.
• The study successfully reconstructed the location of a moving object using hydrodynamic data.

Conclusions:

• Recurrent neural networks, particularly LSTM, are more effective for hydrodynamic object localization using ALL systems.
• The ALL system provides a promising approach for near-field hydrodynamic sensing and object tracking.
• The findings suggest that recurrent connections offer advantages in noisy environments for underwater sensing applications.