An Automated Testing and Classification System For Identifying Defects in Nuclear Steam Generator Tubes Using a Learning Vector Quantization Neural Architecture

摘要

Several different methods exist for the nondestructive evaluation of nuclear steam generator tubes for defect characterization. At present, multifrequency eddy current testing has emerged as the dominant method in industry but requires significant human expertise to accurately identify and classify tube defects. New Mexico State University, (NMSU) was recently funded under the U.S. Department of Energy's joint initiative with NEPO (Nuclear Energy and Power Organization) to design an automated method for the diagnosis and classification of steam generator tube defects. The NMSU research includes a focus on defects characterized as intergranular attack/stress corrosion cracking (IGA/SCC) occurring at tube support plates. The automated detection software developed at NMSU uses a sophisticated set of transform algorithms that provide data visualization capabilities for the raw eddy current data. In addition, the software parameterizes the eddy current data for a second stage defect analysis by an artificial neural network architecture called Learning Vector Quantization (LVQ). This article provides an overview to the automated data capture and visualization program developed at NMSU and specifically focuses on the training issues involved in using neural architectures for defect detection and characterization.