Multi-scale Feature Fusion with Structural Re-parameterization for Lightweight Defect Detection in Power Systems with Embodied Intelligence Perception

Abstract

The reliable detection of defects in power system equipment with embodied intelligent perception is critical for ensuring operational safety and enabling autonomous monitoring. Traditional methods often struggle with low detection accuracy on specific defect types such as small-scale defects, occluded defects, and structural anomalies and are computationally intensive, limiting real-time deployment on edge devices. To address these challenges, we propose MSFSR-LDD, a novel framework that integrates multiscale feature fusion with structural re-parameterization and lightweight optimization. The method first aggregates features across multiple scales to capture both local and global defect patterns. During training, a rich and deep network structure is utilized to learn expressive representations, which is then re-parameterized into a compact structure for efficient inference. Furthermore, channel pruning and knowledge distillation are employed to achieve model compression without significant performance degradation. Experimental results on representative datasets demonstrate that MSFSR-LDD achieves superior defect detection accuracy while significantly reducing model size and inference latency, enabling practical deployment in autonomous and edge-enabled power monitoring scenarios.