基于改进MobileNetV2的钻杆计数方法

张栋; 姜媛媛

doi:10.13272/j.issn.1671-251x.2022060019

基于改进MobileNetV2的钻杆计数方法

doi: 10.13272/j.issn.1671-251x.2022060019

张栋^1,,
姜媛媛^{1, 2, ,}

1.
安徽理工大学电气与信息工程学院, 安徽淮南　232001
2.
安徽理工大学环境友好材料与职业健康研究院(芜湖), 安徽芜湖　241003

基金项目: 安徽省重点研究与开发计划项目(202104g01020012)；安徽理工大学环境友好材料与职业健康研究院研发专项基金项目(ALW2020YF18)。

详细信息

作者简介:
张栋(1997—)，男，安徽淮南人，硕士研究生，主要研究方向为计算机视觉图像处理，E-mail：1023857721@qq.com

通讯作者:
姜媛媛(1982—)，女，安徽颍上人，教授，博士，主要研究方向为人工智能机器学习，E-mail：jyyll672@163.com

中图分类号: TD67
计量
- 文章访问数: 246
- HTML全文浏览量: 55
- PDF下载量: 42
- 被引次数: 0
出版历程
- 收稿日期: 2022-06-07
- 修回日期: 2022-10-07
- 网络出版日期: 2022-09-19

Drill pipe counting method based on improved MobileNetV2

ZHANG Dong^1
,,
JIANG Yuanyuan^{1, 2
, ,}

1.
School of Electrical and Information Engineering, Anhui University of Science and Technology, Huainan 232001, China
2.
Institute of Environment-friendly Materials and Occupational Health, Anhui University of Science and Technology, Wuhu 241003, China

摘要

摘要: 针对现有基于人工及仪器的钻杆计数法存在精度较低、耗时耗力，现有基于图像处理的钻杆计数方法难以提取图像特征，网络模型复杂度高、计算量大等问题，提出了一种基于改进MobileNetV2的钻杆计数方法。通过摄像头采集钻机工作状态图像，采用数据增强对采集的图像进行预处理，在MobileNetV2的基础上，添加卷积注意力模块增强特征的细化能力，优化目标函数提升识别精度，通过迁移学习获取初始参数。将改进后的MobileNetV2作为钻机工作状态识别模型，提取钻机工作状态特征，通过识别钻杆钻进完整过程中装钻杆、打钻杆、卸钻杆、停机4种钻机工作状态生成置信度数据，通过滑动窗口对置信度数据进行滤波，统计钻杆数量，明确钻孔深度。实验结果表明：改进后的MobileNetV2模型识别准确率达99.95%，与经典分类模型ResNet50，Xception，InceptionV3，InceptionResNetV2，MobileNetV2相比，准确率分别提升了1.35%，1.28%，1.43%，0.85%，1.25%，参数量比MobileNetV2模型减少了38.9%，模型收敛速度更快，综合性能更好。将基于改进MobileNetV2的钻杆计数方法应用于煤矿综采工作面的钻杆计数中，平均钻杆计数精度为98.4%，实现了钻杆精确计数，验证了该方法在复杂环境下应用的可行性和实用性。
- 煤与瓦斯突出 /
- 钻机 /
- 钻杆计数 /
- 图像处理 /
- 卷积神经网络 /
- 深度学习 /
- 注意力机制 /
- MobileNetV2
Abstract: The existing drill pipe counting methods based on manual and instrument have the problems of low precision, time-consuming and labor-consuming. The existing drill pipe counting methods based on image processing are difficult to extract image features, the network model has high complexity and large amount of computation. In order to solve the above problems, a drill pipe counting method based on improved MobileNetV2 is proposed. The working state image of the drilling rig is collected through a camera. The collected image is preprocessed by adopting data enhancement. On the basis of MobileNetV2, the convolutional block attention module is added to enhance the thinning capability of features. The objective function is optimized to improve the recognition precision. The initial parameters are obtained through transfer learning. The improved MobileNetV2 is used as the working state recognition model of the drilling rig. The working state features of the drilling rig are extracted by the model. The confidence data are generated by recognizing the four working states of the drilling rig, including drill pipe installation, drill pipe driving, drill pipe unloading and shut down during the whole drilling process of the drill pipe. The confidence data are filtered through a sliding window. The number of drill pipes is accurately counted, and the drilling depth is determined. The experimental results show that the recognition accuracy of the improved MobileNetV2 model reaches 99.95%. Compared with the classical classification models ResNet50, Xception, InceptionV3, InceptionResNetV2 and MobileNetV2, the accuracy is improved by 1.35%, 1.28%, 1.43%, 0.85% and 1.25% respectively. The parameter is reduced by 38.9% compared with the MobileNetV2 model. The convergence speed of the model is faster and the comprehensive performance is better. The drill pipe counting method based on the improved MobileNetV2 is applied to the drill pipe counting of fully mechanized mining face of a coal mine. The average drill pipe statistical precision is 98.4%. The accurate counting of the drill pipes is realized. The feasibility and practicability of application of the method in the complex environment are verified.
- coal and gas outburst /
- drilling rig /
- drill pipe counting /
- image processing /
- convolutional neural network /
- deep learning /
- attention mechanism /
- MobileNetV2

HTML全文

图 1 CBAM结构

Figure 1. Structure of convolutional block attention module

下载: 全尺寸图片幻灯片

图 2 迁移学习框架

Figure 2. Transfer learning framework

下载: 全尺寸图片幻灯片

图 3 基于P−MobileNetV2模型的钻杆计数流程

Figure 3. Drill pipe counting process based on P-MobileNetV2 model

下载: 全尺寸图片幻灯片

图 4 钻机工作状态图像

Figure 4. Images of drilling rig working state

下载: 全尺寸图片幻灯片

图 5 图像增强示例

Figure 5. Image enhancement example

下载: 全尺寸图片幻灯片

图 6 置信度滤波过程

Figure 6. The process of confidence filtering

下载: 全尺寸图片幻灯片

图 7 不同模型的训练曲线

Figure 7. Training curves of different models

下载: 全尺寸图片幻灯片

表 1 P−MobileNetV2模型结构

Table 1. Improved MobileNetV2 model structure

输入大小	步长	卷积核	重复次数
224×224×3	−	CBAM	1
224×224×3	2	Conv 3×3	1
112×112×32	1	Bottleneck	1
112×112×16	2	Bottleneck	2
56×56×24	2	Bottleneck	3
28×28×32	1	Bottleneck	4
28×28×64	2	Bottleneck	3
14×14×96	2	Bottleneck	3
7×7×160	1	Bottleneck	1
7×7×320	1	Conv 1×1	1
7×7×320	−	CBAM	1
7×7×1280	−	AvgPool7×7	1
7×7×1280	1	Conv 1×1×4	1

下载: 导出CSV

表 2 不同模型的训练结果

Table 2. Training results of different models

模型	参数量/byte	训练时间/s	准确率/%
InceptionResNetV2	54 342 884	5 393	99.10
ResNet50	23 595 908	3 027	98.60
InceptionV3	21 810 980	2 333	98.52
Xception	20 869 676	3 922	98.67
MobileNetV2	2 281 348	966	98.70
P−MobileNetV2	1 392 722	1 096	99.95

下载: 导出CSV

表 3 不同钻杆计数方法的统计结果

Table 3. Statistical results of different drill pipe counting methods

方法	实际钻杆数量	检测钻杆数量	准确率/%
人工计数	500	482	96.4
文献[9]	500	485	97.0
文献[10]	500	484	96.8
文献[19]	500	488	97.6
P−MobileNetV2	500	492	98.4

下载: 导出CSV

参考文献(19)

[1]	郑仰峰,翟成,辛海会,等. 煤巷掘进工作面强弱耦合能量控制防治煤与瓦斯突出理论与方法[J]. 采矿与安全工程学报,2021,38(6):1269-1280. ZHENG Yangfeng,ZHAI Cheng,XIN Haihui,et al. Theories and methods of coal and gas outburst prevention by strong-weak coupling energy control in coal roadway driving face[J]. Journal of Mining & Safety Engineering,2021,38(6):1269-1280.
[2]	孙志飞,吴银成,胡云. 钻杆长度测量方法[J]. 工矿自动化,2015,41(3):51-53. SUN Zhifei,WU Yincheng,HU Yun. Length measurement methods of drill pipe[J]. Industry and Mine Automation,2015,41(3):51-53.
[3]	胡少兵,罗明璋,程峰,等. 基于应力波频谱图的护栏金属立柱埋深检测法[J]. 公路,2022,67(6):336-341. HU Shaobing,LUO Mingzhang,CHENG Feng,et al. Method of detecting the buried depth of guardrail metal column based on stress wave spectrum image[J]. Highway,2022,67(6):336-341.
[4]	徐钊,房咪咪,周红伟,等. 基于电驻波的锚杆长度无损测量方法[J]. 工矿自动化,2013,39(9):112-115. XU Zhao,FANG Mimi,ZHOU Hongwei,et al. Non-destructive measurement method of anchor stock length based on electricity standing wave[J]. Industry and Mine Automation,2013,39(9):112-115.
[5]	李宏达,黄鼎琨,张彬,等. 改进的低压脉冲法对变压器绕组变形的探测研究[J]. 南京理工大学学报,2020,44(1):15-20. LI Hongda,HUANG Dingkun,ZHANG Bin,et al. Research in detection of winding transformer variation based on improved LVI method[J]. Journal of Nanjing University of Science and Technology,2020,44(1):15-20.
[6]	徐志强,吕子奇,王卫东,等. 煤矸智能分选的机器视觉识别方法与优化[J]. 煤炭学报,2020,45(6):2207-2216. XU Zhiqiang,LYU Ziqi,WANG Weidong,et al. Machine vision recognition method and optimization for intelligent separation of coal and gangue[J]. Journal of China Coal Society,2020,45(6):2207-2216.
[7]	董立红,彭业勋,符立梅. 基于Sobel边缘检测的圆周Harris角点检测算法[J]. 西安科技大学学报,2019,39(2):374-380. DONG Lihong,PENG Yexun,FU Limei. Circular Harris corner detection algorithm based on Sobel edge detection[J]. Journal of Xi'an University of Science and Technology,2019,39(2):374-380.
[8]	董立红,王杰,厍向阳. 基于改进Camshift算法的钻杆计数方法[J]. 工矿自动化,2015,41(1):71-76. doi: 10.13272/j.issn.1671-251x.2015.01.018 DONG Lihong,WANG Jie,SHE Xiangyang. Drill counting method based on improved Camshift algorithm[J]. Industry and Mine Automation,2015,41(1):71-76. doi: 10.13272/j.issn.1671-251x.2015.01.018
[9]	高瑞,郝乐,刘宝,等. 基于改进ResNet网络的井下钻杆计数方法[J]. 工矿自动化,2020,46(10):32-37. GAO Rui,HAO Le,LIU Bao,et al. Research on underground drill pipe counting method based on improved ResNet network[J]. Industry and Mine Automation,2020,46(10):32-37.
[10]	党伟超,姚远,白尚旺,等. 煤矿探水卸杆动作识别研究[J]. 工矿自动化,2020,46(7):107-112. DANG Weichao,YAO Yuan,BAI Shangwang,et al. Research on unloading drill-rod action identification in coal mine water exploration[J]. Industry and Mine Automation,2020,46(7):107-112.
[11]	SANDLER M, HOWARD A, ZHU M, et al. MobileNetV2: inverted residuals and linear bottlenecks[EB/OL]. [2022-05-16]. http://arxiv.org/pdf/1801.04381.pdf.
[12]	WOO S, PARK J, LEE J Y, et al. CBAM: convolutional block attention module[C]. European Conference on Computer Vision, Munich, 2018: 3-19.
[13]	朱劲松,李欢,王世芳. 基于卷积神经网络和迁移学习的钢桥病害识别[J]. 长安大学学报(自然科学版),2021,41(3):52-63. ZHU Jinsong,LI Huan,WANG Shifang. Defect recognition for steel bridge based on convolutional neural network and transfer learning[J]. Journal of Chang'an University(Natural Science Edition),2021,41(3):52-63.
[14]	LIN T Y,GOYAL P,GIRSHICK R,et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence,2020,42(2):318-327.
[15]	HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, 2016: 770-778.
[16]	CHOLLET F. Xception: deep learning with depthwise separable convolutions [C]. IEEE Conference on Computer Vision and Pattern Recognition, Hawaii, 2017: 1251-1258.
[17]	SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]. IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, 2016: 2818-2826.
[18]	SZEGEDY C, IOFFE S, VANHOUCKE V, et al. Inception-v4: Inception-ResNet and the impact of residual connections on learning [C]. The 31st AAAI Conference on Artificial Intelligence, Palo Alto, 2017: 4278-4284.
[19]	方杰,李振璧,夏亮. 基于ECO−HC的钻杆计数方法[J]. 煤炭技术,2021,40(11):186-189. FANG Jie,LI Zhenbi,XIA Liang. Drill pipe counting method based on ECO-HC[J]. Coal Technology,2021,40(11):186-189.