Abstract:

By using the artificially-fabricated fatigue crack samples as the research objects, the eddy cureent testing （ECT） signals of fatigue crack are collected and are then denoised by means of wavelet transform, with the signal feature being also extracted. Afterwards, a destructive testing procedure is performed to obtain the true profiles of the crack. Based on a parametric model of the fatigue crack, the radial basis function （RBF） neural network is trained with the preprocessed ECT signals and crack shape parameters. Moreover, a great deal of initial populations of crack shape parameters are generated as the inputs of the trained RBF neural network. Thus, the predicted ECT signals, namely the outputs of the network are obtained. An improved genetic strategy is finally applied to the iterative inversion optimization to search the optimal crack shape. Reconstruction results show that the proposed method is of high speed and precision.

Key words: natural crack, eddy current testing, wavelet transform, neural network, forward model, genetic algorithm, shape reconstruction