Learning-Based Approaches to Current Identification from Magnetic Sensors

Area of Science:

• Electrical Engineering
• Computational Electromagnetics
• Applied Mathematics

Background:

• Direct electric current measurement is often infeasible due to accessibility or technical constraints.
• Magnetic field measurements near current sources can estimate currents, but this involves solving an Electromagnetic Inverse Problem (EIP).
• Traditional EIP solutions rely on regularization schemes, while behavioral approaches offer an alternative without strict adherence to physical laws.

Purpose of the Study:

• To systematically investigate the impact of learning parameters in behavioral models for reconstructing Electromagnetic Inverse Problems (EIPs).
• To compare the performance of behavioral approaches against established regularization techniques for linear EIPs.
• To provide a benchmark for evaluating different methodologies in EIP modeling.

Main Methods:

• A benchmark linear Electromagnetic Inverse Problem (EIP) was utilized for practical illustration.
• Classical regularization methods were applied to estimate source currents from magnetic field data.
• Behavioral modeling approaches were employed, analyzing the influence of various learning parameters.
• Analogous correcting actions were developed for behavioral models, mirroring classical regularization techniques.

Main Results:

• Both classical regularization and behavioral models, when appropriately adjusted, can achieve comparable results in solving linear EIPs.
• The study demonstrates that learning parameters in behavioral models play a crucial role in the accuracy of EIP reconstruction.
• Systematic comparison highlights the strengths and weaknesses of both classical and neural network-based approaches for EIPs.

Conclusions:

• Behavioral models offer a viable alternative to classical regularization for solving Electromagnetic Inverse Problems (EIPs), particularly when physical constraints are challenging.
• Careful control and understanding of learning parameters are essential for the successful application of behavioral approaches in EIP reconstruction.
• The findings suggest that neural network-based methods can achieve results similar to traditional techniques, opening avenues for further research in data-driven inverse problem solutions.