Transformer Enhanced based YOLOv8 Integration: A Hybrid Deep Learning Framework for Intelligent Insulator Defect Detection in High-Voltage Transmission Systems

Umer Farooq¹Fan Yang^1*Jamshed Ali Shaikh^2,3*

• ¹State Key Laboratory of Power Transmission Equipment, Systems Security and New Technology, School of Electrical Engineering, Chongqing University, Chongqing 400044, China
• ²Chongqing University, Chongqing, China
• ³School of Computer Science Engineering, South China University of Technology, Guangzhou 510006, China

Insulators are vital components of high-voltage power transmission systems, where undetected defects 9 can lead to catastrophic failures and significant economic losses. Accurate and timely detection of 10 insulator defects (IDs) under diverse environmental conditions is critical for ensuring system 11 reliability. This paper presents Transformer-Enhanced YOLOv8 (TE-YOLOv8), a novel hybrid deep 12 learning framework designed to address the challenges of detecting small, complex defects in 13 transmission line inspections. TE-YOLOv8 integrates transformer-based attention mechanisms with 14 the advanced YOLOv8 architecture, introducing several key innovations that enhance its performance. 15 Specifically, it incorporates Global Convolution (GConv) modules to capture extended spatial context 16 for improved feature extraction, C3f-Global Pooling Fusion (C3f-GPF) modules to amplify 17 discriminative features, and Multiscale Information Fusion (MSIF) modules with learnable weights for 18 adaptive multi-scale detection. Additionally, it utilizes Weighted Feature Information Fusion (WFIF) 19 modules for channel-wise attention to refine feature representation, and a Transformer-enhanced neck 20 architecture to model global dependencies and provide enhanced contextual understanding. To 21 improve localization precision and accelerate convergence, the framework adopts the SCYLLA-IoU 22 (SIoU) loss function. Extensive experimental validation on the IDID and CPLID datasets demonstrates 23 that TE-YOLOv8 achieves mean average precision (mAP) scores of 94.2% and 93.8%, respectively, 24 representing improvements of 4.9% and 5.1% over the baseline YOLOv8, and 1.9% and 2.0% over 25 TE-YOLOV8, while maintaining real-time inference at 82 frames per second. Ablation studies, 26 precision-recall curves, and visualization analyses further confirm the effectiveness of TE-YOLOv8 in 27 detecting insulator defects under challenging operational conditions.