摘要
针对地下电缆管道异常图像样本数量稀少、异常检测难度大等问题,提出一种基于重构误差的地下电缆管道异常检测方法。该方法构建了基于编解码器结构的异常检测模型,将卷积注意力模块(CBAM)嵌入编解码器之间以增强网络的特征提取能力,同时,在编解码器之间使用残差连接,以增强模型对图像细节的重建能力。在训练阶段引入伪随机异常模块,提高模型对异常区域的映射重建能力。实验结果表明,所提方法在公开数据集MVTec上的图像级和与像素级的平均AUROC分别为97.2%和95.9%,在自建数据集WEN_piple上的图像级与像素级的平均AUROC分别为99.9%和97.3%。与其他9种异常检测方法相比所提方法能够有效地对电缆管道内的复杂异常情况进行精确的检测。
Aiming at the problems that the number of anomaly image samples of underground cable pipelines is rare and an-omaly detection is difficult,an anomaly detection method for underground cable pipelines based on reconstruction error is p-roposed.This paper established an anomaly detection model based on the structure ofencoder-decoder,and the convolutionalblock attention module(CBAM)is embedded between encoder-decoder to enhance the feature extraction ability of the netw-ork.At the same time,residual connections are used between encoder-decoder to enhance the ability of the model to reconstruct image details.In the training process,a pseudo-random anomaly module is introduced to improve the mapping reconstr-uction ability of the model for anomaly areas.The experimental results show that,the average AUROC of image-level and pixel-level of the proposed method on the public dataset MVTec is 97.2%and 95.9%,respectively,and the average AUROCof image-level and pixel-level on the self-built dataset WEN_piple is 99.9%and 97.3%,respectively.Compared with other n-ine anomaly detection methods,the proposed method can effectively and accurately detect the complex anomalies in cable p-ipelines.
作者
冯兴明
戴云峰
丁亚杰
马云鹏
李庆武
FENG Xingming;DAI Yunfeng;DING Yajie;MA Yunpeng;LI Qingwu(State Grid Jiangsu Power Supply Co.,Ltd.,Yancheng Power Supply Branch,Yancheng,Jiangsu 224000,China;Changzhou Zhongneng Power Technology Co.,Ltd,Changzhou,Jiangsu 213022,China)
出处
《自动化与仪器仪表》
2023年第9期104-109,共6页
Automation & Instrumentation
基金
国网双创孵化培育资金项目,输配电地下电缆管道验收与测绘机器人系统研制(JF2022015)。
关键词
地下电缆管道
异常检测
卷积注意力模块
编解码器
伪随机异常
underground cable conduit
anomaly detection
convolutional attention module
encoder-decoder
pseudorandom anomaly