基于深度学习的航拍光伏板红外图像热斑检测方法研究

针对光伏电站光伏板热斑故障难以检测的问题，结合无人机巡检技术，提出一种基于深度卷积神经网络的光伏板热斑快速检测方法。首先设计了光伏板识别模型，将Yolov4主干特征提取网络替换成轻量级网络MobileNetV2，并将PAnet网络中标准3×3卷积替换为深度可分离卷积，实现了将光伏板快速从红外图像中识别出来。为快速识别热斑并解决光伏板反光噪声问题，将MobileNetV2网络引入DeeplabV3+模型中，改进由于下采样造成的目标缺失，并将交叉熵损失函数修改为Dice损失函数来进一步提高分割精度。试验结果表明，该方法能够准确识别光伏板热斑，光伏板识别准确率为99. 56%，检测速度为22. 1帧/秒。光伏板识别后的热斑分割准确度达到95. 99%，交并比mIou达到85. 58，检测速度为24. 5帧/秒，该方法能够满足光伏板故障检测的需要。

Photovoltaic hot spot detection of aerial infrared image based on deep learning

In view of the difficulty in detecting hot spots of photovoltaic panels in power stations in China, combined with UAV inspection technology, a fast detection method of hot spots of photovoltaic panels based on deep convolutional neural network was proposed. Firstly, a photovoltaic panel recognition model was designed. The Yolov4 backbone feature extraction network was replaced by the lightweight MobileNetV2 network, and the standard 3×3 convolution in PANet was replaced by the deeply separable convolution, which could realize the rapid recognition of photovoltaic panels from infrared images. In order to quickly identify hot spots and solve the problem of reflective noise of photovoltaic panels, MobileNetV2 network is introduced into deeplabv3 + model, improve the target loss caused by sampling and the cross entropy loss function is modified to dice loss function to further improve the segmentation accuracy. The experimental results show that the method can accurately identify hot spots of photovoltaic panels, with an accuracy of 99. 56% and a detection speed of 22. 1 frames per second. The hot spot segmentation accuracy of photovoltaic panel recognition reaches 95. 99%, MIoU reaches 85. 58%, and the detection speed is 24. 5 frames per second. This method can meet the needs of photovoltaic panel fault detection.