Self-filtering electrical area sensors emerging from deep learning

In this work, we introduce the concept of self-filtering electrically-conductive area sensors for use in nondestructive evaluation, state estimation, and other contemporary engineering applications. For this, a deep learning-based approach is used to optimize sensors' conductivity distributions in order to minimize the effect of noise on Electrical Impedance Tomography (EIT) reconstructions. From sensor design examples, it is shown that (a) the underlying physical self-filtering mechanism arising from the deep learning-based approach functions by reducing the sensitivity of the internal electric potential field to noise perturbations and (b) the use of optimized self-filtering sensors can improve the robustness to noise by approximately 25% compared to standard homogeneous sensors. Lastly, using an alternative interpretation of the self-sensing optimization problem, it is shown that fundamental connections between modeling errors, measurement noise, and physics of self-filtering sensors can be linked and unlocked using deep learning.