A composite neural network model with many sub-network stages is presented.The flaw identification is accomplished stage by stage.Each stage consists of different sub-networks and is trained by a stochastic learning algorithm independently.Numerical simulations of crack identification have shown that the composite neural network model decreases the input vector dimensions of neural network from

N

2

to

N

simplifies the structure of network

saves the learning time

and has a better generalization ability.It can be used effectively in eddy current testing.