Abstract: To address the challenges of weak signal interference and feature extraction in ultraviolet (UV)imaging detection of partial discharges (PD)in 220 kV substation high-voltage equipment,this paper proposes a detection method that partial discharge detection of unmanned aerial vehicle ultraviolet imaging based on multi-scale feature fusion network(MSFF-Net).Firstly,an improved Retinex algorithm and adaptive wavelet threshold denoising are applied to enhance UV images,improving the signal-to-noise ratio in weak discharge regions.Then,a multi-scale feature fusion network is constructed,which combines parallel atrous convolution modules with an attention mechanism to extract and dynamically weight key discharge features under different receptive fields.Experimental results on real-world 220 kV substation datasets demonstrate that the proposed method significantly improves the sensitivity to weak PD signals and successfully identifies over 90% of early-stage discharge defects.The proposed approach effectively enhances the detection and localization capabilities for UV imaging-based PD signals,greatly improving the accuracy and reliability of substation equipment condition monitoring.

Key words: multi-scale feature fusion, UAV-based ultraviolet imaging, attention mechanism, partial discharge detection