Wavelet tools for improving the accuracy of neural network solution of electromagnetic inverse problems

A neural network model is proposed to treat inverse problems in electromagnetics, which includes wavelet functions to improve local approximation capabilities. This processor couples the advantages of an interpretation of the problem based on “features” to the accuracy derived from using wavelets where local corrections are needed. The combined model allows one to cope with singularities of the mapping and to slightly modify the mapping in real time. The detection and characterization of a circular defect in a conducting plate by using eddy current testing is shown to take advantage from the proposed approach in a test case, when unforeseen disturbances are present