基于深度学习的输变电设备紫外放电光斑分割方法

随着紫外成像技术的发展，高压电力设备对于紫外成像图谱的量化分析提出了更高的要求。紫外图谱的量化分析需要用到除紫外成像仪所输出“光子数”额外的紫外光斑图像信息，所以需要将紫外放电光斑从可见光的背景中分割出来。然而，传统紫外图谱光斑分割方法仍存在复杂背景及小光斑分离困难、特征选取复杂、分割精准度低等问题。基于上述问题，提出了一种基于深度学习的紫外图谱光斑分割提取的方法。首先，采用紫外成像仪拍摄电力设备放电缺陷紫外图谱；其次，分别构建FCN-32s、FCN-16s、FCN-8s 3种全卷积网络(fully convolutional networks, FCN)子模型架构，并利用随机梯度下降法进行模型训练；最后，实现输变电设备放电缺陷紫外图谱主光斑的自主分割提取。经过对FCN 3种子模型架构的训练、测试和对比分析，结果表明：FCN-16s模型为紫外光斑分割提取的最佳模型，测试准确率可达99.34%。结果表明基于深度学习的紫外图谱光斑分割方法准确高效，为紫外光斑的量化提取及电力设备放电缺陷的紫外诊断提供了参考。

Transmission and Distribution Equipment Ultraviolet Discharge Spot Segmentation Method Based on Deep Learning

With the development of ultraviolet imaging technology, high-voltage power equipment has put forward higher requirements for the quantitative analysis of ultraviolet imaging spectra. The quantitative analysis of the ultraviolet spectrum needs to use extra ultraviolet spot image information in addition to the "photon number" output by the ultraviolet imager, so the ultraviolet discharge spot needs to be separated from the visible light background. However, the traditional UV spectrum spot segmentation methods still have problems such as complex background and small spot separation difficulties, complex feature selection, and low segmentation accuracy. Based on the above problems, a method for segmentation and extraction of ultraviolet spectra based on deep learning was proposed. First, UV spectra of electrical equipment discharge defects were shot by using UV imager. Then, three FCN sub-model architectures (FCN-32s, FCN-16s, and FCN-8s) were constructed respectively, and the stochastic gradient descent method was used for model training. Finally, the autonomous segmentation and extraction of the main spot of the ultraviolet spectra of the discharge defects of the power transmission and transformation equipment was realized. After training, testing, and comparative analysis of FCN"s three sub-model architecture, it is concluded that the FCN-16s model is the best model for UV spot segmentation and extraction, and the test accuracy rate can reach 99.34%. The results show that the ultraviolet spectroscopy spot segmentation method based on deep learning is accurate and efficient, which provides a reference for the quantitative extraction of ultraviolet spots and the ultraviolet diagnosis of electrical equipment discharge defects.