Training bioinspired sensors to classify flows.

Area of Science:

• Fluid dynamics
• Robotics
• Machine learning

Background:

• Aquatic organisms sense ambient flow signals for navigation and survival.
• Translating biological sensing abilities to underwater robotic vehicles is a key challenge.
• Previous work utilized neural networks with vorticity sensors for flow classification.

Purpose of the Study:

• To systematically investigate the impact of different sensor types on flow pattern classification accuracy.
• To evaluate sensor performance under data corruption.
• To develop instantaneous flow pattern classification using distributed sensor arrays.

Main Methods:

• Trained neural networks to classify vortical flows using four distinct sensor types: vorticity, parallel velocity, transverse velocity, and flow speed.
• Assessed network performance with simulated data corruption.
• Developed and tested a network for instantaneous classification using a spatially-distributed sensor array.

Main Results:

• Networks trained with transverse velocity sensors demonstrated superior accuracy compared to other sensor types.
• Transverse velocity sensors maintained high accuracy even with significant data corruption.
• A network utilizing a small, spatially-distributed sensor array achieved remarkable accuracy in instantaneous flow classification.

Conclusions:

• Transverse velocity sensors are highly effective for flow pattern classification in robotic applications.
• Distributed sensor arrays offer a promising approach for real-time flow pattern recognition.
• This research provides a foundation for developing adaptive sensory systems for unsteady flows.