Flow velocity estimation using a fish-shaped lateral line probe with product-moment correlation features and a neural network

Environmental studies on fish require measurements of highly turbulent flows in both the laboratory and in the field. A fish-shaped bioinspired flow measuring device is applied in conjunction with data processing workflow which leverages the interactions between the body and the surrounding flow field for velocity estimation in turbulent flows. Our objective is to develop a robust velocity estimation methodology relevant for studies of fish behavior using a bioinspired fish-shaped artificial lateral line probe (LLP). We show that the device is capable of covering the range of flow velocities from 0 to 1.5 m/s. Three different sets of experiments performed in a closed flow tunnel, a model vertical slot fishway and laboratory open channel flume were collected and combined to provide time-averaged flow velocity and LLP measurements under fully turbulent flow conditions. Based on the experimental results, a signal processing workflow using Pearson product-moment correlation coefficient (PCC) features in conjunction with an artificial neural network (ANN) is presented. Using PCC features provides a simple data fusion methodology exploiting the use of the LLP's as a simultaneous collocated sensing array. In this work we show that (1) the PCC-ANN workflow provides the first LLP velocity estimator without repeated calibration across the full span of 0-1.5 m/s, (2) using all pressure sensors results in the best performance with R2=0.917, but requires a PCC feature matrix of 55 dimensions and (3) a stepwise reduction of the PCC feature matrix allows for the use of as few as 11 dimensions, and results in R2=0.911, indicating that a modest reduction in LLP velocity estimation performance can be gained by a large reduction in dimensionality. A surprising finding was that after stepwise reduction, the best performing sensor pair combinations were not the expected pitot-like anteroposterior couples spanning from nose to body. Instead, it was found that optimal velocity estimation using the LLP exploited a network of sensor pairs. It is shown that the LLP can be implemented similar to an ADV for highly turbulent flows over the range of 0-1.5 m/s, and in addition provides body-centric pressure distributions which may aid in the interpretation of fish hydrodynamic preferences in future environmental studies.