Robust real-time model parameter extraction in inductive spectroscopy based on Artificial Neural Networks

Abstract

Impedance spectrum modelling requires parameter extraction by solving the corresponding inverse identification problem, for which both gradient-based and stochastic methods have been proposed. Stochastic methods are more robust and have a lower risk of getting trapped in local minima but need a long calculation time. In this study, we examine the implementation of artificial neural networks (ANNs) in solving inverse identification problems in inductance spectroscopy, contrasting their performance with robust stochastic methods. In order to overcome the shortage of a representative amount of experimental data, we propose to use the analytically based model to generate accurate enough labelled data for the training process of the ANN. The artificial data have been structured in a homogeneous and tightly spaced grid in the parameter space, thus supporting the model’s generalisation and suppressing overfitting. ANNs with varied degrees of complexity have been investigated by modifying the number of neurons and evaluated by training and comparison with stochastic parameter extraction methods. The investigation concludes that, for the presented application in inductive spectroscopy, the neural networks can provide comparable parameter extraction results with a relative deviation of 0.03 % of the parameter value and a significant reduction in runtime from 60 s to 8 ms.