MOGAによる最適設計NNを用いた直流電位差欠陥診断法  [in Japanese] DC Potential Non-Destructive Test by Using Neural Network Designed by Multi Objective Genetic Algorithm  [in Japanese]

This paper describes a non-destructive identification of a crack in objects by using DC potential method and a feed forward neural network as a diagnosis method. Neural network for detecting cracks is trained by the fluctuation of potential distribution in the samples of known models. Samples for learning are prepared by simulations using boundary element method. Appropriately studied neural network is able to diagnose unknown cracks immediately. For obtaining neural network which has good reliability of identification, it is necessary to decide optimum parameters such as unit number and coefficient of learning function. Multi objective genetic algorithm is applied as a method of designing the neural network. Multi objective genetic algorithm searches parameters of neural network on the basis of the mean square error at studying and judging. The parameters are selected optimally without heuristic designing. By using the neural networks designed by multi objective genetic algorithm, cracks in objects are estimated with good accuracy.