Swept frequency eddy current material profiling using radial basis function neural networks for inversion

Traditional methods for inverting swept frequency or pulsed eddy current signals to get material information involve iterating with a forward model until the response from the model under the same excitation condition is as close to the measured signal as possible. Although the feasibility of the model based inversion has been demonstrated, the complexity of such procedures and the computational resources that this technique requires has hampered its widespread acceptance in industry (Eaton and Hohmann, 1989). Recent approaches include using the look up tables for features extracted from the signals (Sehuraman and Rose, 1995). The performance of look up table approach depends on the choice of the features extracted. We propose an innovative approach of using a neural network (NN) to solve this inversion problem. Although the use of NN for inverting uniform field eddy current data has been demonstrated, this is the first effort to investigate the feasibility of NN inversion of swept frequency and pulsed eddy current data for thickness measurements of metallic coatings of metal substrates (Udpa and Udpa, 1990; Udpa and Udpa, 1990a). The authors previously reported initial results from this research (Kat ragadda et al., 1996). The current paper focuses on the PC based instrumentation and software developed for the swept frequency material profiler. Results of the NN based classification are summarized, and potential applications discussed.