Physics-Informed Encoder-Decoder Network for EMI Source Reconstruction via Phaseless Single-Plane Near-Field Scanning

-This paper presents a neural-network-based radiation source reconstruction framework designed to simultaneously enhance efficiency and flexibility. An Encoder-Decoder architecture is adopted, in which the decoder incorporates the analytical expression of an equivalent dipole radiation field, constraining the network output to retain explicit physical meaning. A lightweight convolutional feature extraction module is embedded within the encoder to reduce input dimensionality and computational cost. During training, single-plane phaseless near-field data acquired by scanning are fed into the physicsinformed encoder-decoder network (PIEDNet), and the meansquare error (MSE) between the measured and reconstructed fields is minimized to achieve end-to-end estimation of the dipole distribution. Compared with existing neural-network approaches, the proposed method requires only single-plane phaseless data and outputs equivalent dipoles directly, yielding superior generalization; the resulting dipole model can be seamlessly integrated into mainstream electromagnetic interference (EMI) prediction workflows. Ablation studies confirm the effectiveness of the convolutional feature extraction module. Numerical simulations and experimental validations demonstrate that the proposed framework outperforms both the dynamic differential method and traditional machine-learning techniques in reconstruction accuracy and efficiency, offering a new technical avenue for radiation source reconstruction.