This study addresses the challenge of segmenting instances in substations, where complex shapes and occlusions of various devices hinder edge feature extraction and subsequently reduce segmentation accuracy is low. We propose a 3D instance segmentation model based on differential 3D U-Net. First, we introduce a voxel view enhancement module that enhances voxel features and improves the model's spatial perception by integrating a cross-attention mechanism with double view projection. Second, we construct a high-frequency sensitive differential feature fusion module to enhance the model's ability to learn edge features of instances. Finally, we design a deep fusion module to combine deep semantic features with shallow structural features, thereby enhancing the model's semantic discrimination capabilities. Experimental results demonstrate that the proposed model achieves a mean average precision score of 66.58% under a custom power scenario and 70.29% on the ScanNet dataset, significantly outperforming the mainstream models and showcasing its strong engineering application potential.