Evaluation of neural networks for data classification, recognition, and navigation in the marine environment.

摘要

The purpose of this research is to explore a family of strategies using neural networks for automating certain tasks of underwater exploration. Specifically, this research explores the use of neural networks to create basic blocks for detection and classification of a variety of sensor inputs as well as for vehicle control in marine settings. This body of work can then be used to implement fully autonomous detection and navigation systems for use in the marine environment. Several candidate neural network paradigms were evaluated for use in the research and are discussed. This research has been broken into two main portions based on the functional task desired and the nature of the data to be analyzed.; The first project deals with passive acoustic data from hydrophones. In this effort, different preprocessing and network strategies are evaluated for utility in discerning different acoustic sources. Both unsupervised (Kohonen Map) and supervised hybrid paradigms were tested (Kohonen/Multi-Level Perceptron). Sources include surface and underwater vehicles, geophysical sounds, underwater mammals of several types, and several fish species. Recognition rates of 100% are achieved for man made sources and most cetacean sources. Fish are problematic for a combined network, but improved results are achieved using a “fish only” network with pulse energy gathering. With wavelet preprocessing, recognition rates of 31 to 72% are reported for fish only data with 7 species of fish.; The second project examines the case of multiple sensor inputs, including temperature, turbidity, salinity, and pressure. The concept of primitive and emergent behaviors is developed, and the structures are tested on both theoretical and “real world” data sets. A standardized set of common “primitive features” is defined for all sensor types and complex environmental features are recognized as combinations of the primitive features using a multi-level perceptron network. For primitive features in 5% noise, feature extraction of 75.5% is demonstrated. For emergent features, recognitions of 83% are achieved, with some features such as tidal inlets and hydrothermal vents being recognized with 100% correct recognition. Recognition rates as functions of input data format, noise levels, and output category structure are also presented.