Application of improved BP network in the flaws evaluation of conductive materials

The back-propagation (BP) network is widely recognized as a powerful training tool of the multilayer neural networks (MLNNs). Usually it suffers from a slow convergence rate and often results in local minimums, since it applies the steepest descent method to update the network weights. A variety of related algorithms have been introduced to address that problem. Levenberg-Marquardt algorithm is one of the fastest types of these algorithms. This paper presents an approximation calculation for Hesse matrix to train the neural networks when the second order item can not be omitted, and the improved algorithm is successfully used in the surface flaws evaluation of the conductive materials based on eddy current testing (ECT). © 2009 IEEE.