Additive neural network approximation of multidimensional response surfaces for synthesis of eddy-current probes

Abstract

A universal method for constructing multidimensional approximation models of surface eddy current probes with various structures of excitation systems is proposed. This method presupposes, firstly, the description of the eddy-current probe interaction with a testing object. Secondly, a computer design of the computational experiment based on quasi-random sequences with a minimum discrepancy is constructed. Thirdly, an approximation model is created using additive neural network regression and the assessment of the adequacy and information content of the obtained multivariate approximation model. The examples of approximation dependences of the eddy current density distribution are considered. Currents are generated by the uniform surface eddy current probes in the testing object. The results of reproduction of multidimensional response surfaces in the form of level lines are given. The lines were obtained on the basis of approximation models with an assessment of their quality using scatter diagrams. The value of the relative model error in the form of distribution histograms is presented.