Deep Learning Methods for Assessing Time-Variant Nonlinear Signatures in Clutter Echoes

Abstract

The biosonar systems of bats in families of horseshoe and Old-World leaf-nosed bats include peripheral dynamics where the outer ears undergo fast rotations and deformations during echo reception. These motions impart time-variant linear and nonlinear effects on the received echoes. In the present study, we have investigated whether such time-variant effects create discriminable and reliable signatures in clutter, i.e., echoes that are created by a superposition of reflections from multiple, unresolved scatterers. We have used a laboratory setup with artificial foliage that was agitated by fans to create large data sets of clutter echoes. These echoes were triggered by pulses with different time–frequency signatures (constant-frequency, frequency-modulated, and a compound of the two) and received by flexible biomimetic pinna that was actuated via strings to create two different motion shapes at five different motion speeds. Different deep-learning architectures (ResNets, transformers, and a 2D convolutional neural network) were tested for their ability to classify the different motions based on single clutter echoes. The achieved performances (up to 97% overall correct classifications) demonstrated that the time-variant signatures in the clutter echoes could form a reliable substrate for the encoding of sensory information that may provide functional advantages for navigating complex natural environments.