基于深度学习的变电站电力设备可视故障图像识别与诊断研究  

Research on Visual Fault Image Recognition and Diagnosis of Substation Power Equipment Based on Deep Learning

作　　者：赵晓晗 ZHAO Xiao-han(Changyuan Shenrui Relay Protection Automation Co.,Ltd.,Shenzhen,Guangdong 518057)

机构地区：[1]长园深瑞继保自动化有限公司,广东深圳518057

出　　处：《江西电力职业技术学院学报》2024年第10期5-8,共4页Journal of Jiangxi Vocational and Technical College of Electricity

摘　　要：受图像质量影响,利用其进行故障诊断时,往往存在可靠性偏低的问题。为此,进行基于深度学习的变电站电力设备可视故障图像识别与诊断研究。应用最大类间方差阈值分割法进行图像分割处理,提取所需检测电气设备的红外图像后,对电气设备部分进行直方图均衡化,再应用拉普拉斯算法提取图像的高频成分,实现增强处理。在识别诊断阶段,用含旋转角度的矩形框捕捉电气设备在图像中的姿态,引入了深度学习算法中的RetinaNet,在RetinaNet框架的neck部分融合CBAM注意力机制与路径聚合网络(PAN),实现变电站电力设备可视故障图像关键特征的捕捉,在不同尺度上识别具体的故障状态。在测试结果中,不同故障诊断的Acc始终稳定在0.95以上。Conventional visual fault diagnosis methods for substation power equipment often suffer from low reliability due to image quality limitations.To address this challenge,this study proposes a deep learning-based approach for visual fault recognition and diagnosis.First,infrared images of electrical equipment were segmented using the maximum inter-class variance(Otsu)threshold method.The target equipment regions were then enhanced through histogram equalization,followed by high-frequency component extraction via Laplacian algorithm to sharpen critical features.For recognition and diagnosis,rotating bounding boxes were employed to capture equipment spatial orientations.An improved RetinaNet framework was developed by integrating the Convolutional Block Attention Module(CBAM)and Path Aggregation Network(PAN)into its neck structure,enabling multi-scale feature fusion for precise fault state identification.Experimental results demonstrated robust performance with diagnostic accuracy consistently exceeding 0.95 across various fault types.

关 键 词：深度学习 可视故障图像 最大类间方差阈值分割法 直方图均衡化 拉普拉斯算法 RetinaNet CBAM注意力机制 

分 类 号：TM76[电气工程—电力系统及自动化]