Abstract
Analog sensors are widely deployed in electronic systems but are highly susceptible to electromagnetic interference (EMI). Moreover, their electromagnetic susceptibility (EMS) responses, manifested as EMI-induced offsets, vary significantly with temperature. Therefore, accurate modeling of temperature-dependent EMS responses, particularly in sparse temperature–frequency regions with abrupt variations, is critical to ensure stable operation across wide temperature ranges. This letter proposes a kernel-enhanced deep learning approach for modeling analog sensor EMS responses by using sparse measurement data. A deep neural network with temperature, interference frequency, and interference power as inputs is employed to predict output-voltage offsets. A kernel-based data augmentation strategy is employed to enrich critical temperature–frequency regions, and the optimal augmentation intervals are identified through Bayesian optimization. To validate the proposed approach, EMS experiments are conducted on two typical analog sensors by using bulk current injection and direct power injection tests. The results demonstrate that the proposed approach achieves higher prediction accuracy than three baseline methods, especially in temperature–frequency regions exhibiting rapid susceptibility variations.
Funding Information
  • Special Project for High-Quality Development and Special Scientific Research Project of Civil Aircraft (MJZ5-2N22)

This publication has 17 references indexed in Scilit: