Understanding the Flow-signal correlation in seal whisker array sensing using a supervised deep learning model

Phocid seals detect and track artificial or biogenic hydrodynamic trails based on mechanical signals of their whisker arrays. Behavior studies have shown that by using the collective array signals resulting from the interaction with the oncoming vortices, seals are able to locate and distinguish the upstream obstacles. In this study, we investigate the correlation of characteristics of upstream obstacles, flow structures and whisker array signals using a supervised deep learning neural network model. A circular plate is placed in front of a realistic harbor seal head to generate hydrodynamic wake structures; one-way FSI was then simulated to obtain the dynamic behavior and root mechanical signal of each whisker in the two whisker arrays on the seal head. The locations and orientations of the circular plate are systematically varied to generate a large training data set. The network takes the inputs of the temporal mechanical signals of the two arrays and is trained to predict the location and orientation of the plate. The primary features of array signals are identified by incorporating an autoencoder into the network model and are further related to the vortex structures to understand their correlations. This study would provide helpful insight into seal whisker flow sensing mechanisms.