SK-FRCNN: A Fault Detection Method for Hot Spots on Photovoltaic Panels

Photovoltaic power generation is clean and environmentally friendly, and has been widely used. Hot spots on photovoltaic panels, caused by shading and leading to heating, reduce the efficiency of photovoltaic power generation and even damage the panels. To address the problem of low detection accuracy in existing models for hot spot detection on photovoltaic panels, a method for detecting hot spot faults on photovoltaic panels, called SK-FRCNN (Selective Kernel-Faster RCNN), based on the Faster RCNN network is proposed. The feature extraction module in the Faster RCNN network uses an attention mechanism to adaptively adjust the receptive field for hot spots of different sizes, enhancing the feature extraction capability. A feature pyramid structure is adopted with a bottom-up fusion to obtain higher resolution features from low-resolution feature maps through upsampling and fusion operations, thereby preserving multi-scale information of the feature maps and improving the recognition effect for small hot spot faults. The ROI Align unit is used to replace the pooling layer in the Faster RCNN model, canceling the quantization operation and using bilinear interpolation algorithm to obtain floating-point pixel image values, more accurately calculating the feature representation of ROI regions and improving the localization accuracy of hot spot recognition. Experimental results show that the improved algorithm achieves an average detection accuracy of 79.98% for hot spot faults on photovoltaic panels, which is 1.82% higher than that of the original Faster RCNN network.

Automated summary

This paper proposes a method for detecting hot spot faults on photovoltaic panels, called SK-FRCNN (Selective Kernel-Faster RCNN), based on the Faster RCNN network. The method improves the detection accuracy by using an attention mechanism, feature pyramid structure, and ROI Align unit.