Neural networks : an application to electrochemical noise data

Neural networks were applied to the analysis of electrochemical noise data. Electrochemical noise is defined as the fluctuations in either current or potential with time for a metal which is immersed in a conductive solution. This data is of interest because of its relationship to particular corrosion processes. Specifically, a system which is experiencing uniform corrosion will produce a different noise signal than one which is experiencing localized (perforation) corrosion. The economic effects of corrosion are significant and methods which improve the ability to detect, measure and predict corrosion would be extremely valuable.Two series of experiments were conducted. The data for both series were collected from aluminum samples immersed in various aqueous solutions. The series differed from each other in the configuration and programming of the potentiostat which collected the data. The first series only dealt with potential noise while the second series dealt with both potential and current noise. Auxiliary parameters, such as the pH and chloride concentration of the solutions were used in the second series. The first series studied data from only two solutions, while the second series included six solutions.It was possible for neural networks to correctly categorize systems in Series 1 according to the class of corrosion being observed (uniform or perforating). Appropriate data transformation steps were required to effect these classifications and it was also observed that many of these data transformations would lead directly to categorization without the use of a neural network.The additional data collected in Series 2 allowed a more complex analysis. Neural networks were able to simultaneously predict both the propensity towards localized corrosion and the metal dissolution rate. This application demonstrated the power of neural networks.Several types of neural networks and learning algorithms were included in this study. The two systems used most were a backpropagation (multi-layer perceptron) and a radial basis system. Comparisons of the various network systems with regard to speed and accuracy were made.Department of Mathematical SciencesThesis (M.A.