ニューラルネットワークによる渦電流探傷信号の自動認識

概要

論文の詳細を見る
We report the application of the neural network technique to the automatic characterization of defects detected by eddy current testing (ECT) of steam generator tubes of pressurized water reactors. Four parameters are determined to describe the trajectory of the ECT signal, and eight parameters, which come from two trajectories for two different AC current frequencies, are used as the input of the neural network. The orthogonal nature of these parameters is examined by comparing the reproducibility of the learned data by the neural networks using different sets of input parameters. The performance of the neural network is found to be improved by employing the probabilistic descent method for the back-propagation calculation, multiple networks for different defect types, and multiple teaching signals with small noises for one defect to make the network insensitive to noises. Finally, the generalization capability of the network is demonstrated.
一般社団法人日本機械学会の論文
1996-08-25