ULTRASONIC NDE OF ADHERED T-JOINTS USING LAMB WAVES AND INTELLIGENT SIGNAL-PROCESSING

This paper examines the application of artificial neural networks to t he estimation of geometrical parameters of an adhered aluminium T-join t using ultrasonic Lamb waves (s(0)+alpha(1)). Modulus FFTs of receive d signals were applied as inputs to conventional feed-forward networks , which were trained using the delta rule with momentum. The success r ate of various network structures in recognising bond categories was s tudied as a function of the density of information applied to the netw ork inputs and the number of hidden nodes in the network. An optimum n etwork structure appears to exist that will solve a number of problems of this type.