Hydrodynamic Imaging with Artificial Intelligence: detecting submerged objects at a distance using a 2D-sensitive flow sensor array and neural networks

This thesis explores the information that is left behind by underwater objects in their wake, how this information can be measured, and how it can be used to identify properties of these objects, such as their relative location and shape. This application is bio-inspired by the natural ability of fish to measure their fluid flow surroundings and interpret these flows for a range of tasks and behaviors. These tasks include localizing food and detecting obstacles, but this flow sensation is also used for schooling and courtship. To mimic and improve this ability, a novel type of fluid flow sensor is developed, capable of measuring the fluid flow speed in two dimensions for the first time, which is shown to be beneficial for hydrodynamic imaging: determining the properties of moving objects by measuring their produced flow. An array of these flow sensors is deployed at different length-scales, from several centimeters to several meters. The measurements resulting from nearby objects in motion are used to show that hydrodynamic imaging can be scaled up considerably from its biological dimensions. The flow measurements are processed with a variety of artificial neural networks and deep learning methods, including the ELM, MLP, ESN, LSTM, and CNN. These are shown to be well-suited for determining an object’s location, its direction of motion, or its shape.