Machine learning approach for impedance locus uncertainties

Abstract This work deals with the determination of the uncertainty of measurement data, determined by electrical impedance spectroscopy. Four different types of sand were measured impedimetrically in a measuring cell designed as a plate capacitor in a frequency range from 20 Hz to 1 MHz. The measuring cell was filled ten times with each sand and 20 impedance spectra were recorded for each filling. The uncertainty at each frequency was determined from the measurement data. It was found that the measurement data variance with a given measuring-cell filling was negligibly small. However, it increased by a factor of up to 100 when the measuring cell was repeatedly emptied and re-filled with the same material. We propose a way to estimate a continuous approximation of the uncertainty band of the impedance locus in the complex plane from the discrete uncertainties at each frequency. It uses a Support Vector Machine (SVM) to generate a regression curve using the discrete uncertainties. The result of the regression was used to estimate the uncertainties of an average impedance locus. The said machine learning tool can handle large amounts of data, classes, and influencing variables. In this manner, it can help to identify cause-effect relationships. Furthermore, at the end of this work a possibility to estimate a continuous uncertainty band along the impedance locus curve via SVM regression is shown. This is an extension to the common methodology in literature, where the uncertainty is only determined at selected individual points of the impedance spectrum.