«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年08期

页码:

197-201

栏目:

新型计算应用系统

出版日期:

2024-08-10

文章信息/Info

Title:

Research on Indicator Diagram Classification Algorithm Based on Improved ResNet

文章编号:

1673-629X(2024)08-0197-05

作者:

李建平1; 董永杨2; 宋明会3

1. 东北石油大学 环渤海能源研究院,河北 秦皇岛 066004; 2. 东北石油大学 计算机与信息技术学院,黑龙江 大庆 163318; 3. 中国石油集团长城钻探工程有限公司录井公司,辽宁 盘锦 124010

Author(s):

LI Jian-ping1; DONG Yong-yang2; SONG Ming-hui3

1. Bohai Rim Energy Research Institute,Northeast Petroleum University,Qinhuangdao 066004,China; 2. School of Computer and Information Technology,Northeast Petroleum University,Daqing 163318,China; 3. Logging Company of CNPC Great Wall Drilling Engineering

关键词:

抽油机井; 示功图; 深度学习; ResNet; SE子结构

Keywords:

pumping unit well; indicator diagram; deep learning; Residual Network; Squeeze-Excitation substructure

分类号:

TP391.4

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0126

摘要:

示功图是反映抽油机井工作状态的重要图示,通过分析示功图的闭合曲线形状,可以得出抽油机井的具体工作状态,从而可以判断出抽油机井是否发生故障以及具体的故障类型。 随着深度学习的发展,基于深度神经网络的示功图分类也逐渐应用到了抽油机井工况检测当中。 该文提出了基于改进 ResNet 的示功图分类算法,通过优化残差结构和引入 SE 子结构等措施,提高了分类准确性和鲁棒性。 改进的残差结构嵌入了 SE 子结构,对输入特征进行降维的同时也减小了参数的数量,在降低计算量的同时也添加了更多非线性因素,通过不断增加有效特征的权重,不断减小无效特征的权重,进而完成了特征重标定,不仅起到加速网络收敛的作用,也使模型更加轻量化,从而提高了模型的性能。 相较于其它模型,改进的 ResNet 模型可以更好地适应示功图分类任务,分类效果更好。 实验结果表明,基于改进 ResNet 的示功图分类 算法在精确率、召回率和 F1 值上均优于其它示功图分类算法。 该研究为抽油机井工况检测系统提供了更好的理论支持。

Abstract:

The indicator diagram is an important diagram reflecting the working state of the pumping unit well. By analyzing the closed curve shape of the indicator diagram,the specific working state of the pumping unit well can be obtained,so that whether the fault occurs and the specific fault type can be judged. With the development of deep learning,the classification of indicator diagram based on deep neural network has been gradually applied to the condition detection of pumping wells. We propose a classification algorithm of indicator graph based on improved ResNet. By optimizing residual structure and introducing SE substructure, the classification accuracy and robustness are improved. The improved residual structure is embedded with the SE substructure,which reduces the number of parameters while reducing the dimension of input features,and adds more nonlinear factors while reducing the computational load. By continuously increasing the weight of effective features and continuously reducing the weight of invalid features, the feature re - calibration is completed,which not only accelerates network convergence,but also makes the model more lightweight. Thus the performance of the model is improved. Compared with other models, the improved ResNet model can better adapt to the task of indicator diagram classification,and the classification effect is better. The experimental results show that the improved ResNet indicator graph classification algorithm is superior to other indicator graph classification algorithms in terms of accuracy,recall and F1 value. This study provides a better theoretical support for the condition detection system of pumping unit.

常用功能

更新日期/Last Update: 2024-08-10