A Lightweight Multidomain Feature-Enhanced Network for Transformer Oil Leakage Detection in Power Substations

Abstract

Transformer oil leakage detection is essential for ensuring the safe and reliable operation of power equipment in substations. However, complex backgrounds, uneven illumination, reflective surfaces, and the predominance of small-scale leakage regions pose significant challenges to accurate detection. To address these issues, this study proposes MEANet, a lightweight and enhanced object detection framework based on YOLOv11, specifically designed for transformer oil leakage detection. Specifically, a Multidomain Feature Extraction Module (MFEM) is introduced into the backbone to jointly exploit spatial- and frequency-domain information, thereby strengthening multi-scale feature representation. An Efficient Channel-Spatial Attention Module (ECSAM) is incorporated into the neck to suppress background interference and enhance discriminative feature learning. Furthermore, an Adaptive Threshold Focal Loss (ATFL) is designed to alleviate sample imbalance and improve sensitivity to small targets. Experimental results on a transformer oil leakage dataset comprising 1,560 images demonstrate that MEANet achieves 85.6% Precision, 83.2% Recall, and 87.9% mAP@50, outperforming several mainstream detection models while maintaining a lightweight structure (8.2 GFLOPs, 5.4 M parameters). These results indicate that the proposed method provides an accurate and efficient solution for transformer oil leakage detection on the current dataset, suggesting its potential for future integration into intelligent substation inspection systems.