改进YOLOv8的输送带损伤检测方法

doi:10.3969/j.issn.1004-132X.2025.12.003

摘要/Abstract

摘要：

提出一种基于改进YOLOv8算法的输送带损伤检测算法：用Focal Modulation模块替换YOLOv8原有的SPPF模块；针对损伤与背景相似度高的问题，引入DySample轻量动态上采样模块，使采样点集中在目标区域而忽略背景部分，实现损伤的有效识别；在颈部网络中加入高效多尺度注意力模块来获取更多细节信息，进一步提高损伤目标的关注度。引入PIoU v2损失函数，通过计算真实框与预测框之间的重叠面积精准定位损伤，同时考虑长宽比以更好地适应不同形状损伤。实验结果表明，改进后的模型对输送带损伤检测的精确度和平均精确度均值分别达到了90.3%和93.2%，相比于基线模型YOLOv8提高了2.3%和2.5%。改进YOLOv8的检测速度达 83帧/s，可充分满足输送带损伤实时检测的需求。

关键词: 输送带损伤, Focal Modulation模块, 高效多尺度注意力模块, YOLOv8算法, DySample模块, PIoU v2损失函数

Abstract:

A conveyor belt damage detection algorithm was proposed based on improved YOLOv8 algorithm. Focal Modulation module was used to replace the original spatial pyramid pooling-fast（SPPF） module. In view of the high similarity between the damage and the background， a DySample dynamic up-sampling module was introduced to make the sampling points concentrate in the target areas and ignore the background parts， so as to achieve effective damage recognition. The efficient multi-scale attention（EMA） module was added to the neck network to obtain more detailed information and further improve the attention of injury targets. The PIoU v2 loss function was introduced to calculate the overlap areas between the real frame and the predicted frame， and the damage location was more accurate. Considering the aspect ratio might better adapt to the damage of different shapes. Experimental results show that the accuracy and mean average accuracy of the improved model for conveyor belt damage detection reach 90.3% and 93.2% respectively， which are 2.3% and 2.5% higher than the baseline model YOLOv8. The improved detection speed of YOLOv8 algorithm may reach 83 frames /s， which may fully meet the needs of real-time detection of conveyor belt damages.

Key words: belt damage, Focal Modulation module, efficient multi-scale attention module, YOLOv8 algorithm, DySample module, PIoU v2 loss function

中图分类号:

TP391.4

袁媛, 白一超, 周利东, 孟文俊, 王淼, 曲文斌. 改进YOLOv8的输送带损伤检测方法[J]. 中国机械工程, 2025, 36(12): 2829-2836.

Yuan YUAN, Yichao BAI, Lidong ZHOU, Wenjun MENG, Miao WANG, Wenbin QU. Conveyor Belt Damages Detection Based on Improved YOLOv8 Algorithm[J]. China Mechanical Engineering, 2025, 36(12): 2829-2836.

导出引用管理器 EndNote|Ris|BibTeX

链接本文: https://www.cmemo.org.cn/CN/10.3969/j.issn.1004-132X.2025.12.003

https://www.cmemo.org.cn/CN/Y2025/V36/I12/2829

图/表 15

图1 FDEP-YOLOv8网络结构

Fig.1 FDEP-YOLOv8 network structure

图2 实验照片

Fig.2 Experimental photograph

图3 部分增强后的图像

Fig.3 Partially enhanced image

表1 模型训练实验环境

Tab.1 Experimental environment for model training

配置名称	版本/ 参数
深度学习框架	PyTorch 2.3.1+cu118
操作系统	Windows 11
CPU	Intel（R）Core（TM）i7-13650HX CPU@2.60GHz
GPU	NVIDIA GeForce GTX 4060Ti
CUDA	12.4
编译器	Python3.9.19
内存	24 GB

图4 改进模型与相似模型平均精度均值对比

Fig.4 The improved model is compared with the similar model mean average precision

图5 改进模型与相似模型精确度对比

Fig.5 Improved accuracy comparison with similar models

图6 改进模型与基线模型YOLOv8预测效果对比

Fig.6 Comparison of prediction effect between improved model and baseline model YOLOv8

表2 SPPF模块优化对比

Tab.2 SPPF module optimization comparison

模块类型	平均精度P_A/%				精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
模块类型	a	b	c	d	精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
SPPF	90.6	74.0	99.1	99.3	88	90.7	8.1
SPP	91.5	76.0	98.6	99.5	88.9	91.4	7.5
Focal Modulation	86.7	85.0	98.5	99.4	89.3	92.4	8.2
Basicrfb	88.3	82.7	98.8	99.5	91.6	92.3	7.6
SimSPPF	84.4	75.7	98.6	99.5	87.1	89.5	7.5

表3 注意力机制对比实验结果

Tab.3 Comparison of experimental results of attention mechanism

模块类型	平均精度P_A/%				精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
模块类型	a	b	c	d	精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
YOLOv8	90.6	74.0	99.1	99.3	88.0	90.7	8.1
CBAM	88.8	77.8	98.7	99.2	91.2	91.1	8.1
COTA	91.0	75.3	99.1	99.3	87.4	91.2	8.5
Double attention	85.2	82.6	97.1	99.5	87.5	91.1	8.1
EMA	88.2	81.1	98.7	99.5	86.7	91.9	8.1

图7 损失函数边界框损失与分类损失

Fig.7 Loss function boundary frame loss and classification loss

图8 采样模块召回率对比

Fig.8 Comparison of recall rates of sampling modules

表4 消融实验结果

Table 4 Ablation results

基础网络	$a'$	$b'$	$c'$	$d'$	精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
√	×	×	×	×	88.0	90.7	8.1
√	√	×	×	×	86.7	91.9	8.1
√	√	√	×	×	88.6	92.0	8.1
√	√	√	√	×	88.6	93.1	8.2
√	√	√	√	√	90.3	93.2	8.2

表4 消融实验结果

Table 4 Ablation results

基础网络	$a'$	$b'$	$c'$	$d'$	精确度P/%	平均精度均值P_ma/%	计算量/GFLOPS
√	×	×	×	×	88.0	90.7	8.1
√	√	×	×	×	86.7	91.9	8.1
√	√	√	×	×	88.6	92.0	8.1
√	√	√	√	×	88.6	93.1	8.2
√	√	√	√	√	90.3	93.2	8.2

表5 主流算法对比结果

Table 5 Comparison results of mainstream algorithms

算法	平均精度P_A/%				精确度 P/%	平均精度均值 P_ma/%	计算量/ GFLOPS
算法	a	b	c	d	精确度 P/%	平均精度均值 P_ma/%	计算量/ GFLOPS
YOLOV3-tiny	71.1	78.7	92.3	99.3	81.8	85.3	18.9
YOLOV5	86.3	79.3	98.2	99.3	88.3	90.8	7.1
YOLOV6	78.6	82.6	98.6	99.5	85.2	89.8	11.8
YOLOV8	90.6	74	99.1	99.3	88	90.7	8.1
YOLOV8s	89.6	79.9	97.2	99.5	90.6	91.5	28.4
YOLOV9t	83.9	83.3	99.5	99.5	86.2	91.6	7.6
YOLOV9s	92.5	78.2	99.5	99.5	91.7	92.4	26.7
YOLOV10n	87.2	77.2	97.5	98.1	87	90.0	8.2
FDEP-YOLOv8	93.6	80.4	99.4	99.4	90.3	93.2	8.2

图9 Grad-CAM热力图

Fig.9 Grad-CAM thermal map

图10 输送带损伤检测系统界面

Fig.10 Conveyor belt damage detection system interface

参考文献 21

[1]	张梦超，周满山，张媛，等.基于深度学习的矿用输送带损伤检测方法［J］.工矿自动化，2021，47（6）：51-56.
	ZHANG Mengchao， ZHOU Manshan， ZHANG Yuan，et al.Damage Detection Method for Mine Conveyor Belt Based on Deep Learning［J］.Journal of Mine Automation，2021，47（6）：51-56.
[2]	胡璟皓，高妍，张红娟，等.基于深度学习的带式输送机非煤异物识别方法［J］.工矿自动化，2021，47（6）：57-62.
	HU Jinghao， GAO Yan， ZHANG Hongjuan，et al.Research on the Identification Method of Non-coal Foreign Object of Belt Conveyor Based on Deep Learning［J］.Journal of Mine Automation，2021，47（6）：57-62.
[3]	ALVIARI L P， ANGGAMAWARTI M F， SANJIWANI Y， et al. Classification of Impact Damage on a Rubber-textile Conveyor Belt： a Review［J］. International Journal of Mechanical Engineering Technologies and Applications， 2020， 1（1）： 21-27.
[4]	GUO X， LIU X， ZHOU H， et al. Belt Tear Detection for Coal Mining Conveyors［J］. Micromachines， 2022， 13（3）： 449.
[5]	KOZLOWSKI T， WODECKI J， ZIMROZ R， et al. A Diagnostics of Conveyor Belt Splices［J］. Applied Sciences， 2020， 10（18）： 6259.
[6]	熊念强，文锋，李清，等.带式输送机胶带撕裂检测技术综述［J］.煤炭工程，2022，54（7）：79-85.
	XIONG Nianqiang， WEN Feng， LI Qing，et al.Review on Detection Technology of Conveyor Belt Rip［J］. Coal Engineering，2022，54（7）：79-85.
[7]	王以娜.基于视觉检测的皮带纵向撕裂检测关键技术研究［D］. 鞍山：辽宁科技大学，2020.
	WANG Yina，Research on Key Technology of Belt Longitudinal Tear Detection Based on Visual Inspection［D］.Anshan：University of Science and Technology Liaoning，2020.
[8]	陈希阳，李佳楠，张翼，等.基于神经网络的难选夹层磷矿石预分选理论研究［J］.有色金属（选矿部分），2023 （6）：125-131.
	CHEN Xiyang， LI Jianan， ZHANG Yi，et al.Theoretical Study on Pre-separation of Phosphate Ore with Refractory Interlayer Based on Neural Network［J］.Nonferrous Metals（Mineral Processing Section），2023 （6）：125-131.
[9]	樊红卫，刘金鹏，曹现刚，等.低照度尘雾下煤、异物及输送带早期损伤多尺度目标智能检测方法［J］.煤炭学报， 2024，49 （）：1259-1270.
	FAN Hongwei， LIU Jinpeng， CAO Xiangang，et al.Multi-scale Target Intelligent Detection Method for Coal，Foreign Object and Early Damage of Conveyor Belt Surface under Low Illumination and Dust Fog［J］.Journal of China Coal Society， 2024，49 （S2）：1259-1270.
[10]	WANG G， YANG Z， SUN H， et al. AC-SNGAN： Multi-class Data Augmentation for Damage Detection of Conveyor Belt Surface Using Improved ACGAN［J］. Measurement， 2024， 224： 113814.
[11]	GUO X， LIU X， GARDONI P， et al. Machine Vision Based Damage Detection for Conveyor Belt Safety Using Fusion Knowledge Distillation［J］. Alexandria Engineering Journal， 2023， 71： 161-172.
[12]	ZHANG M， SHI H， ZHANG Y， et al. Deep Learning-based Damage Detection of Mining Conveyor Belt［J］. Measurement， 2021， 175： 109130.
[13]	沈景轩，王贡献，孙晖，等.基于图像自适应增强YOLOv5的输送带纵向撕裂检测［J］.仪表技术与传感器，2023（12）：69-74.
	SHEN Jingxuan， WANG Gongxian， SUN Hui， et al.Longitudinal Tear Detection Based on Image-adaptive Enhancement YOLOv5 for Conveyor Belt［J］.Instrument Technique and Sensor，2023（12）：69-74.
[14]	于庆，罗明华，向亮，等.基于改进YOLOv5的矿用输送带纵向撕裂检测方法［J］.矿业安全与环保，2024，51（4）：1-8.
	YU Qing， LUO Minghua， XINAG Liang，et al.Longitudinal Tear Detection Method of Mine Conveyor Belt Based on Improved YOLOv5［J］.Mining Safety & Environmental Protection，2024，51（4）：1-8.
[15]	洪炎，汪磊，苏静明，等.基于改进YOLOv8的煤矿输送带异物检测［J］.工矿自动化，2024，50（6）：61-69.
	HONG Yan， WANG Lei， SU Jingming，et al.Foreign Object Detection of Coal Mine Conveyor Belt Based on Improved YOLOv8［J］.Journal of Mine Automation，2024，50（6）：61-69.
[16]	YANG J， LI C， DAI X， et al. Focal Modulation Networks［J］. Advances in Neural Information Processing Systems， 2022， 35： 4203-4217.
[17]	LIU W， LU H， FU H， et al. Learning to Upsample by Learning to Sample［C］∥Proceedings of the IEEE/CVF International Conference on Computer Vision. Paris ：2023： 6027-6037.
[18]	OUYANG D， HE S， ZHANG G， et al. Efficient Multi-scale Attention Module with Cross-spatial Learning［C］∥ICASSP 2023-2023 IEEE International Conference on Acoustics， Speech and Signal Processing （ICASSP）.Greece： IEEE， 2023： 1-5.
[19]	LIU C， WANG K， LI Q， et al. Powerful-IoU： More Straightforward and Faster Bounding Box Regression Loss with a Nonmonotonic Focusing Mechanism［J］. Neural Networks， 2024， 170： 276-284.
[20]	陈腾杰，李永安，张之好，等.基于改进YOLOv8n+DeepSORT的带式输送机异物检测及计数方法［J］.工矿自动化，2024，50（8）：91-98.
	CHEN Tengjie， LI Yongan， ZHANG Zhihao， et al. Foreign Object Detection and Counting Method for Belt Conveyor Based on Improved YOLOv8n+DeepSORT［J］.Journal of Mine Automation，2024，50（8）：91-98.
[21]	SELVARAJU R R， COGSWELL M， DAS A， et al. Grad-CAM： Visual Explanations from Deep Networks via Gradient-based Localization［J］. International Journal of Computer Vision， 2020， 128： 336-359.