Photovoltaic Module Fault Detection Based on TH-YOLOv11s

Due to factors such as the dense arrangement of photovoltaic modules and complex image backgrounds, issues like information loss, false positives, and missed detections are common in the fault detection process. To address these challenges, this paper proposes an improved TH-YOLOv11s algorithm for photovoltaic module fault detection. First, the original SPPF module is structurally optimized by introducing a triplet attention mechanism and a convolutional residual structure, which enhances the model’s ability to perceive multi-scale fault regions and anisotropic boundary features. The HFERB module is integrated into the C3K2 backbone network, and its high-frequency-enhanced residual structure strengthens the model’s ability to extract high-frequency detail features, effectively improving detection accuracy. Experimental results show that the TH-YOLOv11s algorithm significantly outperforms the original YOLOv11s model across multiple metrics: precision increases by $3.3 \%$, reaching $85.9 \%$; recall improves by $3.1 \%$, reaching $86.9 \%$; [email protected] increases from $85.7 \%$ to $87.8 \%$, an improvement of $2.1 \%$; [email protected]:0.95 increases by $1.8 \%$, reaching $61.4 \%$; and the detection speed reaches 160.2 frames per second, fully meeting the practical requirements of photovoltaic module fault detection.