Abstract: To address the limitations of traditional resistivity inversion methods in mining scenarios—including initial model dependency, boundary ambiguity, and artifacts present in existing deep learning-base inversion approaches—this study proposes a physics-constrained U-Net inversion method. By integrating electrical sensitivity characteristics and depth focusing mechanisms, the method constructs a weighted cross-entropy loss function based on U-Net's multi-scale feature fusion architecture. Enhanced encoder-decoder skip connections are employed to amplify resistivity contrasts between anomalies and background fields. A parameter space for resistivity distribution was defined based on three types of typical anomalous bodies, and forward modeling was performed on 6 000 models using the finite element method. Dipole-dipole array configurations were applied to acquire apparent resistivity profiles, establishing a geoelectric model-response paired dataset for supervised training. Experimental results demonstrate a Dice coefficient of 0.950±0.018 and a reduction in inversion time from 65.2 s (least-squares method) to 1.0 s per instance, improving computational efficiency by 98.5%. The synergistic optimization of physical priors and deep learning provides an effective solution for precise detection of hidden water-conducting structures in coal mine hazard prevention.