基于异质图像信息深度学习的绝缘子污秽状态智能监测

The accurate measurement of contamination status has crucial importance to the prevention of pollution flashover and the enhancement of power system reliability. In this project, deep learning is adopted to utilize heterogeneous information from visible, infrared and ultraviolet images. And then, thermal-electrical coupling dynamic model of wet polluted insulator is used in conjunction with deep learning to realize the intelligent monitoring of contamination grades. Firstly, using visible camera, infrared imager and ultraviolet imager, contaminated insulators in operation are photographed and the heterogeneous image database is obtained. Meanwhile, image segmentation method and specification for shooting are studied. Secondly, Sparse-DBNs is adopted to carry out the deep learning of heterogeneous information from visible, infrared and ultraviolet images and realize the non-contact and online monitoring of contamination grades. Parameters of the Sparse-DBNs such as learning algorithm, sparse penalty function and initialization method are optimized. Thirdly, dynamic model of the whole flashover process is established. The mechanism of interaction between temperature field and electric field is researched. Variation rules of parameters of temperature field and electric field are quantitative studied. After that, thermal-electrical coupling dynamic model of wet polluted insulator in different environment is built. Finally, based on the theoretical research of pollution flashover and heterogeneous image analysis, time-space variations of thermo-electric field are researched, and in conjunction with heterogeneous information deep learning, the.intelligent prediction of pollution flashover is implemented.

绝缘子污秽状态的准确监测对于预防污闪事故发生，提高电力系统可靠性具有重要意义。本项目采用深度学习理论综合利用可见光、红外、紫外三种图像异质信息，结合湿污绝缘子热-电耦合动态模型，实现绝缘子污秽状态智能监测。首先，采用可见光相机、红外热像仪、紫外成像仪对运行中绝缘子进行拍摄，得到不同污秽状态绝缘子的异质图像库，研究适合不同类型图像的分割处理方法。其次，采用稀疏深信度网络对可见光、红外、紫外图像异质信息进行深度学习，并从学习算法、稀疏惩罚函数、初始化方法等方面对其进行优化，实现污秽等级识别。然后，建立绝缘子污闪全过程动态模型，明晰温度场与电场相互作用机理，研究温度场与电场各项参数变化规律，构建不同环境下湿污绝缘子热-电耦合动态模型。最后，将污闪理论研究与异质图像分析相结合，研究不同污闪发展阶段绝缘子表面温度场、电场分布的时空变化规律，结合异质信息深度学习实现污秽闪络的智能预测。

绝缘子污秽状态的准确监测对于预防污闪事故发生，提高电力系统可靠性具有重要意义。课题采用深度学习理论综合利用可见光、红外、紫外三种图像异质信息，结合湿污绝缘子热-电耦合模型及污秽绝缘子泄漏电流特性研究，实现污秽状态的智能监测。进行了高电压试验和变电站内现场作业，采集不同湿度各污秽等级绝缘子表面泄漏电流信号、可见光图像、红外图像、紫外放电视频，建立了泄漏电流及异质图像数据库。开展了污秽绝缘子表面泄漏电流特性研究，利用4项3阶Nuttall窗三谱线插值法对基波及主要谐波的参数进行校正，研究了泄漏电流时域和频域参量在不同环境湿度下随污秽等级的演化特性，提出了一种利用卷积神经网络对泄漏电流吸引子相图进行分类的污秽状态识别方法。依据绝缘子形状特征、污层发热和传热学理论，建立绝缘子热平衡方程，运用勒让德方程导出热平衡方程的通解。基于混合物吸湿特性分析，揭示了污秽中水溶性混合物临界湿度低于任何单一水溶性成分临界湿度这一规律是导致人工污秽试验结果与自然积污绝缘子污闪特性存在差异的重要原因。通过污层吸湿特性和水分蒸发速率分析，研究污层干燥带产生规律。运用有限元分析软件，对污秽绝缘子表面电流密度分布、传热与温度场分布进行热电耦合仿真研究，研究了电流密度分布与温度分布的关系、干燥带形成规律及其与温度分布的耦合作用。研究了污秽绝缘子红外图像及表面温度分布演化规律、紫外图像光斑面积与拍摄参数间关系等问题。采用稀疏自编码器（SAE）和深信度网络（DBN）分别对紫外光斑面积序列和可见光图像色彩特征进行处理，实现了污秽状态识别。将可见光图像的色彩特征、红外图像的温度特征、紫外图像的放电强度特征送入深度收缩自编码器（DCAE），得到异质图像融合特征，利用深度Q网络判断绝缘子的安全状态，实现污秽闪络预警。基于上述研究成果，完成了一套外绝缘污秽状态监测软件系统。