Argentum Solutions Inc. - Sterling Guidance on Corrosion and Materials Degradation

Argentum Solutions, Inc.

Sterling guidance on corrosion and materials degradation

Potential-pH Diagrams

THERMEXPERT - Potential-pH diagram generator

Intelligent Tools

POLEXPERT - Polarization Scan Artificial Neural Network Expert System

SEQEXPERT - Sequential Immersion Test Artificial Neural Network Expert System

CYLEXPERT - Rotating Cylinder Electrode Intelligent Rotation Rate Calculator

Corrosion Calculator

Corrosion Economics Estimator

FINCALCULATOR - Corrosion Economic Calculator

TUTORIAL ON ARTIFICIAL NEURAL NETWORKS

David C. Silverman

Table of Contents

Overview of Tutorial
Artificial Neural Network Background
The Back-propagation Computing Element
The Back-propagation Artificial Neural Network
Training the Back-Propagation Neural Network

Example of Back-propagation Artificial Neural Network
Radial Basis Function Artificial Neural Network
Probabilistic Artificial Neural Network
General Regression Artificial Neural Network
Modular Artificial Neural Network

Overview of Tutorial

Artificial intelligence tends to emphasize rational action and activity surrounding artificial intelligence tends to focus on designing technologies for such action to result in rational decisions. For the most part, imitating the human brain has not been a significant part of that activity. Artificial neural network technology while being classified within the realm of artificial intelligence departs from that approach. Some of the rationale for developing artificial neural networks was to develop computer systems that could imitate some of the functionality of the human brain. An excellent book that discusses artificial intelligence and where artificial neural networks lie within the realm of artificial intelligence is "Artificial Intelligence, A Modern Approach", S. J. Russell and P. Novig, Prentice Hall, 1996. This more focused technology has attributes that makes it applicable to certain aspects of corrosion reasoning and decision making.

Corrosion is a material degradation process resulting from a complex interaction among intrinsic material properties and extrinsic physical and chemical variables in the environment. Small changes in environmental variables and small changes in physical properties can result in large changes in type and rate of corrosion. Great strides have been made in the ability to understand and predict the complex interaction in many systems. This understanding can result in broadly applicable formal rules. But, the corrosion practitioner is often faced with having to understand and predict corrosion in situations in which the knowledge base can only account for some of the effects and formal rules cannot be invoked. Experiments may provide only limited additional information. The result is that the corrosion practitioner knows that a relationship (or pattern) exists among the variables and the resulting degradation but the rules of the relationship cannot be verbalized. The practitioner has recognized a pattern. The ultimate materials decision is made from that pattern and the human experience that recognizes the pattern. That experience, unfortunately, has been departing the corporate scene. Realizing that patterns are used to fill gaps where formal rules and mechanistic knowledge are lacking has led to the exploration of using artificial neural networks to aid in corrosion prediction and to at least partially compensate for that lost experience.

This tutorial has several objectives:

to provide some background on artificial neural networks - what they are and what they may be able to predict.
to summarize structure and proper training of the back-propagation neural network since this structure is the one most often used in corrosion classification applications.
to summarize a number of alternative neural network structures useful for learning classification types of problems. These alternative neural network structures discussed are radial basis functions, probabilistic (Bayesian) neural networks, general regression neural networks, and modular neural networks.
to show by way of a simple example used throughout the tutorial the ability of each of the technologies to classify the same problem. The goal is to show that neural network technology extends well beyond the back-propagation neural network. Indeed, these other neural network structures might be better suited for some classification problems common in corrosion.

Two working examples of back-propagation neural networks tied to expert systems can be accessed elsewhere on this website http://www.argentumsolutions.com. The first predicts field (plant) observation of possible types of corrosion from a laboratory generated potentiodynamic polarization scan in a system called POLEXPERT. The second predicts chemical compatibility in the field (plant) of certain non-metallics from a type of laboratory sequential immersion test in a system called SEQEXPERT. Whereas a number of networks have been reported to tie laboatory data to laboratory observations, these combined systems remain two of the very few that have tied laboratory test results to field (plant) observations. These intelligent systems have also been described elsewhere (A. L. Silverman and D. C. Silverman,"Easily Accessible Intelligent Corrosion Tools on the Internet", Paper 05060, Corrosion/2005, Houston, TX, 2005¹

(596k), E. M. Rosen and D. C. Silverman, "Corrosion Predictions from Polarization Scans Using an Artificial Neural Network Integrated with an Expert System", Corrosion, Vol. 48, No. 9 p. 734, 1992¹ (774k), and D. C. Silverman, "Artificial Neural Network Predictions of Degradation of Nonmetallic Lining Materials from Laboratory Tests", Corrosion, Vol. 50, p. 411, 1994 ¹

(514k)

Next Page: Artificial Neural Network Background

Return to Table of Contents

¹ © NACE International publication and year shown in citation above. All rights reserved. Displayed with permission from NACE International, Houston, TX (http://www.nace.org). Published in Corrosion, in the month and year shown in the citation above.

David C. Silverman, Ph.D. - Primary Consultant
E-Mail:     dcsilverman@argentumsolutions.com
Phone:     314-576-3586
Fax:         314-754-9825
Address:   The Argentum House
                14314 Strawbridge Ct.
                Chesterfield, MO 63017