To achieve millimeter-level 3D reconstruction of complex electrical equipment in substations

this paper proposes a 3D Gaussian Splatting modeling method that integrates feature enhancement with dynamic resolution control. The supervised contrastive learning and structural enhancement modules are introduced into the method

significantly improving feature recognition capabilities in weakly textured and repetitive structural regions. Meanwhile

an adaptive resolution control strategy is designed in the training phase

enabling efficient unified modeling of global structure and local details. Through comparative experiments with multi-view stereo (MVS)

neural radiance field (NeRF)

and the original 3D Gaussian splatting (3DGS) on a self-built multi-view visible light image dataset

the proposed method demonstrates significant advantages in reconstruction metrics including peak signal-to-noise ratio (PSNR)

structural similarity index measure (SSIM)

and L1 norm. The average L1 error is reduced to 0.033

enabling precise restoration of millimeter-level structural details and exhibiting strong engineering applicability.