Silicon-based ultraviolet sensors enable direct and non-contact monitoring of true discharge by capturing ultraviolet emission patterns. This monitoring is of significant value for the electrical safety of industrial facilities, such as textile production lines, where discharge events can lead to equipment failure or safety hazards. However, discharge targets in silicon-based UV images are often weak and small, and they are easily confused with background noise. To address these challenges, this paper proposes PDDA, a discharge detection method built on YOLOv8. PDDA introduces an EMA attention module into the bottleneck to enhance pixel-level feature weighting, replaces the original neck with a Cross-level Context Fusion Module to strengthen multi-scale fusion, and adopts the Swish activation function to improve training stability. Experimental results demonstrate that PDDA achieves 86.9% mAP@0.5 with a throughput of 12 fps, outperforming several mainstream detectors while maintaining practical efficiency for direct, continuous monitoring.

silicon-based ultraviolet sensor, true discharge, object detection, multi-scale feature fusion, indus-trial safety monitoring