Comparison of radial basis function and back-error propagation neural networks for crack detection in variable cross-section beams