Shape Classification Using a Single Seal-Whisker-Style Sensor Based on the Neural Network Method

Area of Science:

• Biomimetics and Sensor Technology
• Hydrodynamics and Fluid Mechanics
• Artificial Intelligence and Machine Learning

Background:

• Aquatic animals like seals use whiskers for target identification and tracking.
• This biological mechanism offers inspiration for developing advanced, low-power, portable, and eco-friendly sensors.
• Existing sensing technologies may lack the sensitivity and adaptability of biological systems.

Purpose of the Study:

• To design and test a seal-whisker-like cylindrical sensor for detecting underwater targets.
• To train and evaluate a convolutional neural network (CNN) using force signals from the sensor.
• To determine the effectiveness of this biomimetic approach in object identification and analyze key signal features.

Main Methods:

• Fabrication of a single seal-whisker-mimicking cylinder.
• Experimental measurement of forces (lift and drag) on the cylinder with nine different upstream targets.
• Development and testing of a convolutional neural network (CNN) model using collected force signal datasets.
• Application of Fourier analysis to understand signal characteristics and model performance.

Main Results:

• The seal-whisker sensor combined with a CNN successfully identified underwater objects in most test cases.
• Certain targets caused confusion, indicating limitations in the current model.
• Increasing signal sample length improved accuracy but did not fully resolve target confusion.
• High frequencies (>5 Hz) were found to be irrelevant for the CNN model's performance.
• Lift signals provided more distinguishing features than drag signals for target differentiation.
• Model efficacy correlated strongly with spectral feature discrepancies in lift signals.

Conclusions:

• A biomimetic sensor inspired by seal whiskers, coupled with a CNN, shows significant potential for underwater object identification.
• Lift signal analysis, particularly its spectral features, is crucial for distinguishing between different targets.
• Further research may be needed to address confusions with specific targets and optimize sensor performance.