A LATERAL LINE INSPIRED APPROACH TO UNDERWATER SENSING WITH A PRESSURE SENSOR ARRAY

Open Access
Author:
Andes, Ryan Christopher
Area of Honors:
Mechanical Engineering
Degree:
Bachelor of Science
Document Type:
Thesis
Thesis Supervisors:
  • Bo Cheng, Thesis Supervisor
  • Hosam Kadry Fathy, Honors Advisor
Keywords:
  • Supervised Learning
  • Neural Networks
  • sensor array
  • bioinspired
Abstract:
Autonomous underwater vehicles (AUVs) are taking on more industry missions due to their ability to travel to remote environments, present low-risk to human operators, and operate continuously with low-power consumption. These vehicles require low-volume, integrated, robust sensory systems that can sense the surrounding flow field and detect underwater objects. Murky and cluttered environments present significant resolution challenges to traditional sonar and optical sensing methods, whose high energy consumption, bulkiness, and active sensing methods are ill-suited for the design and mission requirements of AUVs. Fish possess lateral line sensory systems that allow them to navigate environments, avoid obstacles, and detect prey and predators through an array of surface and sub-surface sensors that perceive fluid motion and pressure gradients. In this experiment, a functional array of pressure sensors was used to extract information from the fluid flow. Information encoded in the unique unsteady wake patterns produced by traversing objects produced a time-dependent pressure distribution acting on the surface of the array. Sequential data from the sensor excitation profiles, derived from these wake disturbances, were inputted into a Long Short Term Memory Recurrent Neural Network capable of learning long-term dependencies in sequential non-linear patterns. The network was then trained to classify an underwater stimulus of unknown geometry. The results indicate that the classification scheme of the LSTM is successful in converting the information encoded in the stimulus wake into a prediction of the stimulus geometry. Sizing the hidden layer around 10 units is shown to result in a high classification accuracy. To verify the classification scheme, a set of test data will need to be created to verify the results.