A machine learning approach for analyzing complex vibrational signals from real seal whiskers

Seals use their whiskers to extract detailed information from the fluid environment and sense prey. We examined the vibration features that correlate with hydrodynamic excitation, using single point laser vibrometery to record the vibrations from harbor seal whiskers in a recirculating water tunnel. The recorded spectra reveal a complex vibrational response with features dependent on flow speed (0.5-2.5 m/s), whisker angle of attack, and presence or absence of an upstream disturbance. A machine learning approach was used to successfully predict the presence of an upstream cylinder, based on the power spectral density (PSD) of the vibrational signal, with comparable performance (>85%) across a variety of simple models. When classifiers were run with amplitude values removed but frequency bin information retained, only a 5% loss in test accuracy resulted, indicating that the relevant feature difference between signals is shape of the PSD, rather than overall amplitudes. In tandem with the experiments, computational fluid dynamics are performed based off of CT scans of seal whiskers. The results from the vibration analysis will be used to inform fluid structure interaction (FSI) computations as a means to connect the fluid dynamics to the frequency response of the whisker.