Presentation Information
[B-4-40]Contactless Wideband Impedance Estimation with Inductively Coupled Magnetic Probes via Machine Learning
〇Souma Jinno1 (1. Osaka Institute of Technology)
Keywords:
Impedance estimation,Inductively coupled magnetic probe,Machine learning
Impedance measurement using inductively coupled (magnetic) probes can estimate the impedance of a device under test without breaking the wiring, and is widely used in EMC applications such as conducted-noise evaluation of power converters. The conventional approach calibrates the probe with known reference standards at every measurement and recovers the impedance by matrix inversion; however, this is an ill-conditioned inverse problem whose accuracy depends strongly on the choice of standards and degrades at high frequencies. In this work, we reformulate the inverse problem as a per-frequency-point supervised regression. Using a programmable impedance board that automatically generates many known impedances by combinatorial switching of passive elements, we systematically acquire a large set of calibration data and train a multilayer perceptron (MLP) that maps the probe's S-parameters and frequency to the impedance. From four boards we collected 1020 patterns over 9 kHz–100 MHz and performed a leakage-free evaluation. The proposed method tracks the ground truth (VNA measurement) over a wide band and is more stable and accurate than the conventional two-standard ABCD calibration. Its key feature is that it replaces the per-measurement physical-standard calibration with a one-time, upfront training.