摘要
煤炭是重要的国家能源,在煤炭的生产、转运的过程中主要依靠输送带传送。然而借助输送带转运时,大块煤炭影响输送带的安全运行。针对煤炭运输过程中的块煤监测问题,提出一种基于轻量化神经网络的输送带块煤实时监测方法(AHE Mosaic Ghost CBAM YOLOv5,AMGC YOLOv5)。首先,利用自适应直方图均衡化(Adaptive Histogram Equalization,AHE)进行图像数据预处理,降低煤尘、粉尘以及煤矿井下光线对目标监测的影响,提高数据集的清晰度和对比度,同时结合Mosaic多数据增强提高数据集的丰富度。其次,引入Ghost Net轻量化神经网络,充分利用特征提取与特征图冗余特性之间的关系,将传统卷积操作与轻量化线性操作相结合,在保证检测精度的同时,极大程度减少模型的参数量和浮点计算量。最后,结合CBAM(Convolutional Block Attention Module)注意力机制依次完成通道特征提取和空间特征提取,改善特征提取的倾向性,提高模型的表征能力,降低背景图像的干扰。试验结果表明:改进以后的AMGC YOLOv5算法在模型轻量化方面,网络层数减少145层,模型体积压缩58.8 M;参数量和浮点计算量分别减少约63.60%和68.86%;训练时长减少约0.76 h,GPU使用率减少约21.78%。此外,在模型性能方面,精确度提升约0.91%,实时监测效率从68.34 FPS提高至94.34 FPS,提高约38.05%。由此可见,AMGC YOLOv5不仅实现轻量化,而且能够有效提升模型的各项检测性能。
Coal is an essential national energy source,which is mainly conveyed by the conveyor belts in the process of coal production and transportation.However,when transferred by means of the conveyor belt,lump of coal affects the safe operation of the conveyor belt.For the problem of lump coal monitoring during coal transportation,so the AMGC YOLOv5(AHE Mosaic Ghost CBAM YOLOv5)lightweight neural network based method for the real time monitoring of lump coal on conveyor belt is proposed.First,Image data pre-processing using adaptive histogram equalization(AHE),reducing the impact of coal dust,dust and underground coal mine light on target monitoring,and to enhance the clarity and contrast of the data set.At the same time combined with Mosaic data to increase the richness of the data set.Subsequently,The Ghost Net lightweight neural network is introduced to make full use of the relationship between feature extraction and redundancy characteristics of feature maps,and combine traditional convolutional operation with lightweight linear operation,which greatly reduces the number of parameters of the model and the amount of floating point computation while ensuring the detection precision.Eventually,combined with Convolutional Block Attention Module(CBAM)mechanism to complete the channel feature extraction and spatial feature extraction sequentially.To enhance the feature extraction process,improve the ability to characterize,and reduce interference from background images.The experimental results showed that:The improved AMGC YOLOv5 algorithm in terms of model lightweight,The number of network layers is reduced by 145,and the model volume is compressed by 58.8 M.The number of parameters and floating point calculations are cut by about 63.60%and 68.86%.Training hours were reduced by about 0.76 hours,and GPU usage was reduced by about 21.78%.In addition,in terms of model performance,Precision was improved by about 0.91%,and real-time monitoring efficiency was improved by about 38.05%from 68.34 FPS to 94.34 FPS.It can be seen that AMGC YOLOv5 not only realizes lightweight,but also can effectively improve the model's various detection performance.
作者
吴利刚
陈乐
吕媛媛
张喆
WU Ligang;CHEN Le;LYU Yuanyuan;ZHANG Zhe(College of Mechanical and Electrical Engineering,Shanxi Datong University,Datong 037003,China;College of Coal Engineering,Shanxi Datong University,Datong 037003,China)
出处
《煤炭科学技术》
EI
CAS
CSCD
北大核心
2023年第S02期285-293,共9页
Coal Science and Technology
基金
2022山西省基础研究计划资助项目(202303021211330)
2022年山西省高等学校科技创新计划创新平台资助项目(2022P009)
关键词
深度学习
神经网络
轻量化模型
注意力机制
块煤实时监测
deep learning
neural Networks
lightweight model
attention mechanism
real-time lump coal monitoring
作者简介
吴利刚(1986—),男,山西大同人,副教授,硕士生导师。E-mail:ligangwu@yeah.net
