Ill-posed problems in electromagnetics: advantages of neuro-fuzzy approaches

This paper aims to show how and why a hybrid neuro-fuzzy data processing approach yields a novel, efficient way to treat ill-posed inverse problems. Two practical examples of such problems in applied computational electromagnetics are presented. The first one concerns the eddy current testing of conducting cylindrical structures in which the use of fuzzy knowledge is shown to improve the capability of discriminating within buried and surface cracks. Secondly, an identification problem in a nuclear fusion application is adequately solved by a priori resolving conflicting goals of the optimization procedure. In both examples, the fuzzy part of the system is basically used to manage the strategy of selection of the proper region of the working space. In this way, the accuracy of the identification is strongly improved. This suggest that the combined use of the fuzzy expansions and of the multidimensional feature extraction capabilities of neural networks can play a relevant role in inverse problem analysis.