Advances in neural network non-destructive testing

A neural system architecture based on a concept of task decomposition, and capable of handling fuzzy variables, aimed to treat nondestructive testing applications is described. To demonstrate the efficiency of this approach an electrostatic sample problem is presented and adequately solved. However, the techniques discussed are of general applicability, and then suitable for practical NDT problems. Special emphasis is placed on the analysis of the manner in which a simple neural network extracts useful information about the problem via the learning process. Starting from this analysis a strategy of decomposition is derived and successfully applied to the test problem. One of the key features of the proposed model is that the performance of the whole system can easily be improved by locating the problematic portions, whereas an ordinary neural network model acts as a black box