Enhanced YOLOv11n-Based Object Detection for Oil Leakage in Substations

This paper presents an enhanced YOLOv11n-based object detection model for identifying oil leakage defects in substations. First, a DCNv4 deformable convolution is integrated into the C3k2 backbone module to overcome the limitations of fixed-kernel convolutions, improving detection of tiny oil droplets and complex backgrounds. Next, a Context-Anchored Attention (CAA) mechanism is embedded into the High-level Selective Feature Pyramid Network (HSFPN) for multi-scale feature fusion, forming a new neck called CAA-HSFPN. Finally, the standard CIoU loss is replaced with an Inner-MPDIoU loss, accelerating convergence and improving small-target localization. Extensive experiments on a self-constructed dataset of 8,844 images show that the proposed method achieves a $3.2 \%$ increase in mAP@50 while reducing model parameters by $25.1 \%$. The results demonstrate a favorable trade-off between detection accuracy, computational efficiency, and deployment feasibility, highlighting the model’s potential for real-world power grid maintenance applications.