This paper describes the work undertaken to de-skill the complex procedure of determining corrosion mechanisms derived from electrochemical noise data. The use of neural networks is discussed and applied to the real time generated electrochemical noise data files with the purpose of determining characteristics particular to individual types of corrosion mechanisms. The electrochemical noise signals can have a wide dynamic range and various methods of raw data pre- processing prior to neural network analysis were investigated. Normalized data were ultimately used as input to the final network analysis. Various network schemes were designed, trained and tested. Factors such as the network learning schedule and network design were considered before a final network was implemented to achieve a solution. Neural networks trained using general and localized corrosion data from various material environment systems were used to analyze data from simulated nuclear waste tank environments with favorable results.
Electrochemical Noise (EN) is no longer a new corrosion monitoring technique ~d commercial/industrial applications have been increasing for the last several years [1- 19]. As such, plant personnel are coming to terms with the usefulness of the technique. One major issue with EN is data interpretation [20]. Although EN based corrosion monitoring is gaining wider acceptance within the corrosion monitoring and plant operation communities, a more succinct method of analyzing and interpreting the results of EN data is needed [14-19]. At present, interpretation of EN data requires detailed analysis by corrosion experts to identify and categorize the mechanism of corrosion attack from the data files. The introduction of a Neural Network method of analysis of the raw EN data to provide an easily understandable interface will help to broaden the range of users able to understand the significance and importance of corrosion data [21].
One of the most notable difficulties for the users of EN as a corrosion monitoring tool is the vast amount of raw data that must be regularly screened and processed. A method of quickly screening raw data and summarizing it in a few pertinent indicators to convey the maximum amount of information is needed, particularly in multiple channel plant applications. If applied correctly, Neural Networks could de-skill the post-data collection procedure and greatly increase the utility of EN as a corrosion monitoring method. In addition, a simplified method of analysis could result in a substantial extension of plant life by giving advance warning of damaging corrosion problems and result in significant cost savings. Neural Networks would achieve this by minimizing the need for on-site or off-site skilled analysis by corrosion experts. In addition, Neural Networks would provide a consistent methodology for data analysis that would be less subjective and time consuming as compared to analyses by corrosion experts [20].
BACKGROUND ON NEURAL NETWORKS
The basic principle of neural networks is to use a set of simple processing elements (neurons) linked either randomly or specifically into a network architecture capable of solving complex computational problems. The main application of these networks at present takes the form of recognizing and subsequently categorizing, completing or modifying input data patterns in some way.
The memory of a neural network is unlike that of conventional computer programs. It takes the form of a threshold value for each of the neurons, along with a weight for each of the inputs to that neuron. The basic principle of the neuron is that the input value be multiplied by the weight value for that input. In t
