Session: 01-06-01: New Advances in Acoustics and Vibration: AI and Machine Learning, New Methods and Materials

Paper Number: 150211

150211 - Artificial Echolocation: Leveraging Neural Networks for Ultrasound-Based Shape Recognition 

Echolocating animals, such as bats and dolphins, are adept at identifying objects using only ultrasound pulses. This remarkable ability is crucial for hunting prey and navigating in their environment, and it has inspired research efforts to replicate these perceptual skills in engineered systems. However, despite advancements in various areas, including broadband and directive acoustic beams, array configurations, and advanced signal processing algorithms, artificial systems have largely failed to match the perception efficiency and accuracy of echolocating animals.

In this presentation, we discuss a novel ultrasound perception framework, which employs multiple shallow convolutional neural networks (CNNs) working in parallel. Each CNN is specialized in recognizing a distinct shape based solely on the acoustic echoes that shape generates. This specialized approach allows us to develop perception algorithms for shape identification that require less training data compared to existing methods. We demonstrate that these algorithms can be easily scaled to learn new objects by simply adding a parallel CNN for each new object, without modifying the already-trained networks. Importantly, we also show that CNNs trained exclusively on synthetic data can accurately classify objects based on real echoes. This is achieved through a combination of shape-specific CNN architecture and physics-informed data augmentation. To simulate three-dimensional acoustic scattering from an object, we use a custom-built numerical model involving the Green’s function method and k-wave, a commercial finite-difference-time-domain solver of the wave equation. The echoes obtained through these simulations are augmented to make the model robust to variations in amplitudes, phase, and noise, and then used to train the CNNs. The trained models are subsequently tested on echoes measured in echolocation experiments.

As a demonstration of our approach, we focus on a few fundamental shapes including sphere, cylinder, and cube, with the size of each object being of the order of the ultrasound wavelength. At these small object sizes, generated echoes are perceptually similar presenting a challenging classification scenario for CNNs. Despite these challenges, our approach shows a remarkable ability to classify these shapes in a physical setting, demonstrating robustness to hardware variability, source uncertainties, experimental noise, and variations in medium acoustic properties. Overall, results demonstrate that these models are scalable, independently customizable, computationally cost-effective, and data-efficient while ensuring accuracy in both synthetic and physical environments. This framework not only mirrors the capabilities of biological echolocation, which are fast, efficient, and sequential learning for perception but also offers practical benefits for imaging and navigation technologies. By bypassing the need for extensive real-world data collection, our approach significantly reduces the cost and effort involved, paving the way for advanced human-made sonar technologies for applications spanning from biomedical diagnosis to autonomous vehicles.

Presenting Author: Ganesh U. Patil University of Michigan Ann Arbor

Presenting Author Biography: Dr. Ganesh U. Patil is a Postdoctoral Research Fellow in the Department of Mechanical Engineering at the University of Michigan Ann Arbor. He received his Ph.D. and M.S. in Mechanical Engineering from the University of Illinois Urbana Champaign, and his Bachelor and Diploma in Mechanical Engineering from VJTI, Mumbai, India. Dr. Patil studies how waves propagate through solids and fluids, and use this understanding to develop new materials and imaging methods to make engineering systems more reliable, efficient, and safer.

Authors:

Ganesh U. Patil University of Michigan Ann Arbor
Hyung-Suk Kwon University of Michigan Ann Arbor
Bogdan I. Epureanu University of Michigan Ann Arbor
Bogdan-Ioan Popa University of Michigan Ann Arbor