程宏波, 王林, 吴浩,等.电力设备红外援例诊断方法研究及其实现[J].河南理工大学学报（自然科学版）,2025,44(3):130-137.

CHENG H B, WANG L, WU H,et al. Research on power equipment case aid diagnosis system based on deep learning[J]. Journal of Henan Polytechnic University(Natural Science) , 2025, 44(3): 130-137.

电力设备红外援例诊断方法研究及其实现

程宏波1,2, 王林1, 吴浩2, 谢子宁1, 李昊岭1

1.华东交通大学 电气与自动化工程学院，江西 南昌  330013；2.人工智能四川省重点实验室，四川 宜宾  643000

摘要: 目的 为发挥红外诊断专家的经验知识，为电气设备故障的诊断及处理提供依据和参考，  方法 提出一种电力设备红外图像援例诊断方法，收集并整理电气设备的典型故障案例，建立电气设备典型故障的红外案例库；通过优化网络全连接层数量，利用全局平均池化代替最大池化对VGG-16深度学习网络进行改进，减少红外图像处理时中间特征的数量，降低图像匹配的计算工作量；利用改进的VGG-16网络提取红外图像的典型特征，计算待测图像与典型案例图像特征之间的余弦距离，利用余弦距离衡量待测图像与标准图像之间的相似程度，筛选最接近的相似案例，以相似案例为依据提供故障原因分析和处理措施建议。收集7 种设备43 类故障的典型案例200 个，开发了电力设备红外援例诊断程序。  结果 实验结果表明，采用改进后的深度学习网络，单张红外图像匹配平均仅需要0.255 s，比SURF方法用时缩短85.5%，比SIFT方法用时缩短91.9%，提取特征进行匹配后诊断正确率可达94.74%。  结论 所提方法可提高红外图像处理的效率，具有较高的诊断准确率，可为电力设备红外诊断提供一种新的方法，诊断结果可整合既有的专家经验知识为现场故障处理提供指导，具有较好的应用潜力。

关键词:深度学习;红外检测;故障案例;援例诊断

doi: 10.16186/j.cnki.1673-9787.2023040029

基金项目:国家自然科学基金资助项目（51967007）；江西省重点研发计划项目（20202BBEL53008）；人工智能四川省重点实验室开放课题（2022RZY01）

收稿日期:2023/04/14

修回日期:2023/09/28

出版日期:2025-04-18

Research on power equipment case aid diagnosis system based on deep learning

CHENG Hongbo1,2, WANG Lin1, WU Hao2, XIE Zining1, LI Haoling1

1.School of Electrical and Automation Engineering， East China Jiaotong University， Nanchang  330013， Jiangxi， China；2.Sichuan Key Laboratory of Artificial Intelligence， Yibin  643000， Sichuan， China

Abstract: Objectives In order to leverage the experience and knowledge of infrared diagnostic experts and provide a basis and reference for the diagnosis and treatment of electrical equipment faults. Methods a method for infrared image assisted diagnosis of power equipment was proposed. Typical fault cases of electrical equipment were collected and organized， and an infrared case library of typical faults of electrical equipment was established. By optimizing the number of fully connected layers in the network and using global average pooling instead of max pooling to improve the VGG-16 deep learning network， the number of intermediate features in infrared image processing was reduced， thereby reducing the computational workload of image matching. The improved VGG-16 network was used to extract typical features of infrared images， the cosine distance between the features of the test image and the typical case image was calculated， the similarity between the test image and the standard image was measured by their cosine distance， the closest similar case was selected to provide fault cause analysis and processing suggestions. 200 typical cases of 43 types of faults in 7 types of equipment were collected and an infrared assisted diagnosis program for power equipment was developed. Results The experimental results showed that using the improved deep learning network， single infrared image matching only took an average of 0.255 seconds， which was 85.5% shorter than that of the SURF method and 91.9% shorter than that of the SIFT method. After extracting features for matching， the diagnostic accuracy could reach 94.74%. Conclusions The proposed method could improve the efficiency of infrared image processing with high diagnostic accuracy， providing a new method for infrared diagnosis of power equipment. The diagnostic results could integrate existing expert experience and knowledge to provide guidance for on-site fault handling， which made it have good application potential.

Key words: deep learning; infrared detection; fault case; aid case diagnosis