基于改进YOLOv8n的立井刚性罐道接头错位检测算法  

An improved yolov8n based dislocation detection algorithm for shaft rigid tank channel joint

作　　者：王满利 杨爽 张长森 WANG Manli;YANG Shuang;ZHANG Changsen(School of Physics&Electronic Information Engineering,Henan Polytechnic University,Jiaozuo 454000,China)

机构地区：[1]河南理工大学物理与电子信息学院,河南焦作454000

出　　处：《煤炭科学技术》2024年第S2期236-248,共13页Coal Science and Technology

基　　金：国家自然科学基金资助项目(52074305);河南省科技攻关资助项目(242102221006)。

摘　　要：立井刚性罐道作为提升系统的导向装置,是提升系统安全稳定运行的基础,使用过程中容易产生横向移位和变形,一旦出现不平顺,将引起提升系统强烈振动,严重影响提升系统运行安全性。为及时发现立井刚性罐道接头故障,消除提升系统运行隐患,提出了一种改进YOLOv8n的立井刚性罐道接头错位检测算法(YOLOv8n-CFW)。首先,针对井筒模糊、黑暗、强光及存在复杂背景成像环境下,YOLOv8n对立井刚性罐道图像特征提取能力不足的问题,融入卷积块的注意力模块(CBAM),该模块将通道注意力机制与空间注意力机制结合,形成一种新的卷积块结构,从而更好地进行特征融合,并且有助于模型更加集中地关注输入图像的重要部分,从而提高模型的识别准确性和泛化能力,克服了YOLOv8n主干网络在模糊、黑暗、强光环境下对局部重要信息特征提取能力不足,泛化能力较差等问题;接着,为轻量化网络,使用Faster_Block代替Bottleneck,降低C2f模块的计算复杂度,克服了YOLOv8n网络的模型参数量过大部署困难的不足;然后,为抑制低质量图像产生的有害梯度,使用WIoU损失函数代替CIoU,WIoU损失函数,通过动态分配梯度增益,抑制低质量图像产生的有害梯度,克服了YOLOv8n网络定位性能不足,使网络在接头错位检测中更加精确地定位到刚性罐道边缘,从而减小误差,并进一步提高模型识别准确性和泛化能力;最后,在推理过程中设置兴趣区域,进一步抑制背景图像干扰,并采用非定焦测距法计算罐道接头偏移尺寸。试验表明,相较于基础YOLOv8n网络,YOLOv8n-CFW检测网络在立井刚性罐道接头数据集上,精度P提升了1.4%,召回率R提升了8.2%,平均精度mAP由YOLOv8n的83.8%提升为90.3%,提高了6.5%,模型大小减少了1.4MB,相比其他YOLO算法,在立井刚性罐道接头错位检测中YOLOv8n-CFW具有显著的优势。As the guiding device of the hoisting system,the shaft rigid tank channel is the basis for the safe and stable operation of the hoisting system,which is prone to lateral displacement and deformation during the use process,and will cause strong vibration of the hoisting system in the event of unevenness,which seriously affects the safety of the operation of the hoisting system.In order to timely discover the faults of vertical well rigid tank channel joints and eliminate the hidden dangers of the hoisting system operation,a vertical well rigid tank channel joint misalignment detection algorithm YOLOv8n-CFW is proposed to improve YOLOv8n.First of all,for the blurred wellbore,darkness,strong light and the existence of complex background imaging environment,the YOLOv8n is insufficient to extract the image features of the rigid tank channel of the vertical well,and it incorporates the CBAM Attention Mechanism Module,which combines the channel attention mechanism with the spatial attention mechanism to form a new convolutional block structure for better feature fusion and helps the model to focus more on the important parts of the input image,thus improving the model's recognition accuracy and generalization ability,and overcoming the problem of YOLOv8n's trunk network's lack of feature extraction ability and generalization ability for locally important information under blurred,dark,and bright light environments.feature extraction ability is insufficient and the generalization ability is poor;then,to lighten the network,Faster_Block is used instead of Bottleneck to reduce the computational complexity of the C2f module,which overcomes the shortcomings of YOLOv8n network that the amount of model parameters is too large to be deployed with difficulty;and then,to inhibit the harmful gradients generated by low-quality images,the WIoU loss function is used instead of the CIoU to improve the anchor frame prediction accuracy,and the WIoU v3 loss function,which suppresses the harmful gradients generated by low-quality images b

关 键 词：刚性罐道 接头错位 YOLOv8n CBAM注意力机制 改进C2f模块 

分 类 号：TN911.73[电子电信—通信与信息系统]