In the intelligent inspection of power systems, the detection of equipment defects is confronted with problems such as low background discrimination, multi-scale morphological differences, and the difficulty in identifying small targets and fine-grained defects, which makes it hard for existing models to balance detection accuracy and computational efficiency. To address this, this study proposes an improved lightweight detection framework, GRID-YOLO. This framework enhances the semantic discrimination ability of the backbone network for complex defects by introducing a cross-stage hierarchical multi-cognitive spatial attention module (C2MSA), designs an enhanced multi-scale bidirectional feature pyramid network (EMFPN) to achieve efficient alignment and preservation of cross-level features, and builds a hybrid attention reparameterization fusion module (HAFusion) to dynamically filter and enhance heterogeneous features. Additionally, a grouped distribution focal loss head (GDFLHead) is adopted to strengthen the collaborative optimization between classification and regression tasks. In this study, a dedicated dataset for substation equipment defect detection was constructed, and extensive comparative experiments were conducted. The results show that GRID-YOLO achieves an mAP50 of 78.3%, which is 6.1 percentage points higher than the baseline model. The mAP50–95, Recall, and F1-score reach 52.3%, 73.6%, and 77.6%, respectively, and the inference speed reaches 506 FPS. Compared with various advanced detection models, GRID-YOLO demonstrates strong overall performance and can better meet the actual needs of substation equipment defect detection in the intelligent inspection scenarios of power systems.