基于ECA-CNN的高压隔离开关状态识别方法
张咪,鲍哲,吴泽锋,江钰韬
(西安西电开关电气有限公司,陕西 西安 710077)
摘 要 :为了解决高压隔离开关状态识别困难、实时监测能力不足的问题,构建了融合嵌入式感知和轻量化神经网络的智能监测体系。硬件层采用模块化设计,智能感知单元与后台控制系统形成闭环联动,捕捉开关动作触头状态关键特征。软件层提出双阶段优化策略:在数据预处理阶段,构建融合几何 变换 ( 旋转 / 剪裁 )、直方图均衡化、图像滤波和噪声注入等构成的图像特征增强算法,有效提升图像 质量并扩充样本多样性;在特征提取阶段,设计深度优化的融合高效通道注意力 (ECA) 机制的卷积神经 网络 (CNN) 模型,通过高效通道注意力机制模块实现通道特征权重动态分配,强化关键特征提取能力。 算法分析验证表明,设计的图像增强算法能够有效提升图像质量,突出关键特征,使识别模型准确率提升超30%;提出的ECA-CNN模型能够在图像增强算法的基础上,进一步提升对图像关键特征的关注度, 其识别准确率高达 97% 以上,较CNN模型提升12%。
关键词 : 高压隔离开关 ;状态识别 ;图像增强 ;卷积神经网络 ;注意力机制
中图分类号 :TM564.1 文献标识码 :A 文章编号 :1007-3175(2025)12-0063-08
A State Recognition Method for High-Voltage Disconnect Switches Based on ECA-CNN
ZHANG Mi, BAO Zhe, WU Ze-feng, JIANG Yu-tao
(Xi’an XD Switchgear Electric Co., Ltd, Xi’an 710077, China)
Abstract: To overcome the challenges of state recognition and insufficient real-time monitoring capability in high-voltage disconnect switches, this study establishes an intelligent monitoring system integrating embedded sensing and lightweight neural networks. The hardware layer adopts modular design, where intelligent sensing units and background control systems form closed-loop linkage to capture critical characteristics of contact states during switching operations. The software layer proposes a dual-stage optimization strategy: In the data preprocessing stage, an image enhancement algorithm combining geometric transformations (rotation/cropping), histogram equalization, image filtering, and noise injection is developed to significantly improve image quality and augment sample diversity. During the feature extraction phase, a deep-optimized convolutional neural network(CNN) model incorporating an efficient channel attention (ECA)mechanism was designed. The efficient channel attention mechanism module enables dynamic allocation of channel feature weights, thereby enhancing the extraction of key features. Algorithm analysis and verification show that the designed image enhancement algorithm can effectively improve image quality, highlight key features, and increase the accuracy of the recognition model by over 30%. The proposed ECA-CNN model can further enhance the focus on key features of the image on the basis of the image enhancement algorithm, with a recognition accuracy rate of over 97%, which is 12% higher than that of the CNN model.
Key words: high-voltage disconnect switch; state recognition; image enhancement; convolutional neural network; attention mechanism
参考文献
[1] 蒋健,陈宇昇,蔡润庆,等 . 电网高压隔离开关机械故障诊断 [J]. 沈阳工业大学学报,2023,45(2):127-132.
[2] 刘伟杰,赵军亮 . 一起隔离开关传动机构失效原因分析及改进 [J]. 高压电器,2024,60(8):269-274.
[3] 公多虎,詹仲强,葛志杰,等 . 基于小波分析的高压隔离开关故障检测方法研究 [J]. 自动化技术与应用, 2024,43(3):19-22.
[4] 关为民,卞超,李东风 . 基于主轴扭矩-转角特性 的高压隔离开关缺陷分析仿真 [J]. 高压电器,2024, 60(1):48-54.
[5] 王旭鹏,郝建,钟尧,等 . 基于 ILPSO-LSSVM的GIS设备隔离开关机械异响振动识别方法研究 [J]. 高压电器,2023,59(11):172-183.
[6] 顾晓东,唐丹宏,黄晓华 . 基于深度学习的电网巡检图像缺陷检测与识别 [J]. 电力系统保护与控制, 2021,49(5) :91-97.
[7] 袁伟,赵玉凯,李杨,等 . 应用于光纤的高压隔离开关状态识别技术 [J]. 传感器与微系统,2024, 43(8):160-164.
[8] 杜耀文,谢静,刘志坚,等 . 基于深度学习的高压隔离开关分合状态检测算法研究 [J]. 电力系统保护与控 制,2023,51(19) :114-123.
[9] 刘子英,张靖,邓芳明 . 基于 B P 神经网络的高压隔离开关分合闸监测识别 [J]. 电力系统保护与控制, 2020,48(5):134-140.
[10] 张靖,单长吉,周丽,等 . 基于图像融合的高压隔离开关分合闸状态识别 [J]. 红外技术,2024,46(5):539-547.
[11] 陈富国,蔡杰,李中旗 . 基于机器视觉的高压隔离开关设备状态判别与故障诊断技术 [J]. 微型电脑应用, 2022,38(2):191-194.
[12] 万嘉龙,况立群,曹亚明,等 . 复合主干融合的图像增强算法 [J]. 计算机系统应用,2024,33(4):194-201.
[13] WANG Qilong, WU Banggu, ZHU Pengfei, et al.ECANet: Efficient channel attention for deep convolutional neural networks[C]//2020 IEEE/ CVF Conference on Computer Vision and Pattern Recognition,2020 :11534-11542.
[14] 马群,赵美蓉,郑叶龙,等 . 基于自适应条件直方图均衡的红外图像细节增强算法 [J]. 红外技术,2024, 46(1):52-60.
[15] 石野,王晓凯,刘慧锋 . 基于局部滤波模板的双边滤波算法研究 [J]. 测试技术学报,2021,35(1):49-53.