Machine learning methods for reconstructing the acoustic fields of bat biosonar

Abstract

Understanding the relationships between biosonar and flight in bats requires synchronized recordings of the echo inputs and the flight kinematics of the animals. Here, integrated arrays of 50 high-speed camera and 32 ultrasonic microphones have been set up to collect the data that is necessary for achieving this goal. The arrays are set up to record bats as they fly through an an obstacle course inside a cylindrical tunnel. Due to the complexity of the received signals in these scenarios, we have been developing custom strategies that rely on a combination of compressed sensing and deep learning to reconstruct the acoustic field inside the tunnel from the microphone-array measurements. By using a combination of the high-speed video and acoustic data, it can be attempted to determine the bat's location in the array from the image data and then reconstruct the properties of the sound fields that were emitted and received at this position using the acoustic array data. The goal of this research is to eventually determine the dynamic characteristics of the bat's biosonar emissions such as beampatterns and potentially time-variant characteristics as well as the biosonar inputs that the bats rely on for controlling their flight behaviors.