欢迎访问《图学学报》 分享到：
图学学报

图学学报 ›› 2022, Vol. 43 ›› Issue (5): 825-831.DOI: 10.11996/JG.j.2095-302X.2022050825

• 图像处理与计算机视觉 • 上一篇    下一篇

基于小样本不均衡数据的供水管道泄漏智能检测算法

  

  1. 1. 广东电力发展股份有限公司，广东 广州 510630；  2. 广东能源集团科学技术研究院有限公司，广东 广州 510630；  3. 华南农业大学图书馆，广东 广州 510642
  • 出版日期:2022-10-31 发布日期:2022-10-28
  • 基金资助:
    国家自然科学基金项目(51775116)；广东能源集团重点科技项目(YJY/20-033) 

Water supply pipeline leakage intelligent detection algorithm based on small and unbalanced data 

  1. 1. Guangdong Electric Power Development Co. Ltd, Guangzhou Guangdong 510630, China;  2. Guangdong Energy Group Science and Technology Research Institute Co. Ltd, Guangzhou Guangdong 510630, China;  3. Library of South China Agricultural University, Guangzhou Guangdong 510642, China
  • Online:2022-10-31 Published:2022-10-28
  • Supported by:
    National Natural Science Foundation of China (51775116); Key Science and Technology Projects of Guangdong Energy Group (YJY/20-033)

PDF (PC)

87

摘要：

针对能源电厂供水管道泄漏视觉检测存在数据样本少、不均衡等问题，提出一种基于小样本不 均衡数据的供水管道泄漏智能检测算法。首先，提出一种基于多掩码混合 Multi-mask mix 的数据增强方法，通 过随机生成掩码层对原始图像进行区域提取与混合，在 Multi-mask mix 中引入支持向量机(SVM)获取管道正常 和泄漏特征，为混合掩码块提供更准确的先验标签；其次，提出一种均衡化策略并应用于图像层面和掩码层面， 以实现数据均衡化；最后，基于深度学习的 Resnet18 网络模型实现管道泄漏检测与识别。实验结果表明，该算 法处理图像数据后可使 Resnet18 模型对管道泄漏识别准确率提升 1.1% ~ 4.4%，说明深度学习模型能有效提升 管道泄漏检测的分类精度，优于现有其他算法。此外，该算法现已成功应用于能源电厂供水管道泄漏检测。

关键词: 小样本, 多掩码混合, 数据增强, 数据均衡化, 管道泄漏检测

Abstract:

To address the problems of few and unbalanced data samples in the visual detection of water supply pipeline leakage in energy power plants, an intelligent detection algorithm for water supply pipeline leakage based on small sample unbalanced data was proposed. First, a data enhancement method based on Multi-mask mix was proposed. The original image was extracted and mixed by the mask layer randomly generated, and the support vector machine (SVM) was incorporated into Multi-mask mix to obtain pipeline normal and leakage features, thus providing more accurate prior labels for the hybrid mask blocks. Secondly, an equalization strategy was proposed and applied to the image level and mask level to achieve data equalization. Finally, a deep learning-based Resnet18 network model was utilized to attain pipeline leak detection and identification. The experimental results show that the algorithm can improve the accuracy of the Resnet18 model for pipeline leakage detection by 1.1%–4.4% after processing image data, and can effectively enhance the classification accuracy of the deep learning model for pipeline leakage detection, outperforming other existing algorithms. In addition, the algorithm has now been successfully applied to the leakage detection of water supply pipelines in energy power plants. 

Key words:  , small sample, Multi-mask mix, data enhancement, data equalization, pipeline leakage detection

中图分类号: