Hydrodynamic object identification using artificial neural networks

Passive object sensing has evolved in aquatic animals to enable them to recognize hydrodynamic objects. This unique capability can be used in autonomous underwater vehicles to gain better awareness of the marine environment. Here, we present a data-driven model that uses artificial neural networks to identify the shape of an obstacle placed in potential flow using data from a stationary sensor array. Specifically, the machine learning framework is used to solve the inverse problem of estimating the body shape from the measured velocity field. The ability of neural networks to deduce the complex underlying relationships without explicit mathematical description is used for parametric fitting of velocity flow data to accurately predict object shape characteristics. Synaptic weights obtained using a gradient-descent based optimization are used to obtain relations between the shape coefficients and the velocity field. Large data sets corresponding to flows with varying object shapes are generated and used to train and validate the performance of this machine learning approach. Finally, this data-driven method is easy to train owing to the analytical nature of the forward problem and is found to accurately estimate object shapes from limited and localized data acquired at a distance.