基于轻量化YOLOv4的发电机定子表面缺陷检测算法

Generator Stator Surface Defect Detection Algorithm Based on Lightweight YOLOv4

目前大型发电机定子表面缺陷检测主要以抽转子的人工检测为主,存在检测周期长,准确率差等问题,论文提出一种基于轻量化YOLOv4的发电机定子表面缺陷检测算法,以腔内爬壁机器人为载体进行定子缺陷检测。将改进的MobileNetv3作为算法的主干特征提取网络,通过在特征融合层引入CSP结构,融合卷积层和BN层的方法,使得论文算法模型体积较YOLOv4大幅减小。实验结果表明,论文算法在本文发电机定子表面缺陷数据集上的平均检测精度为98.3%,优于原始YOLOv4,模型体积比YOLOv4缩小了84.5%,检测速度提高了45.4%,表明了该方法在嵌入式平台上进行发电机定子缺陷实时检测的应用前景。

At present,the surface defect detection of large generator stator mainly relies on manual detection of the extraction rotor,which has problems such as long detection period and poor accuracy.In this paper,the generator stator surface defect detec⁃tion algorithm based on lightweight YOLOv4 is proposed,and the cavity wall-climbing robot is used as the carrier for stator defect detection.The improved MobileNetv3 is used as the backbone feature extraction network of the algorithm.Through the introduction of CSP structure in the feature fusion layer and the fusion of convolutional layer and BN layer,the volume of the algorithm model in this paper is greatly reduced compared with that of YOLOv4.Experimental results show that the average detection accuracy of the pro⁃posed algorithm on the generator stator surface defect data set is 98.3%,which is better than the original YOLOv4,the model vol⁃ume is reduced by 84.5%compared to YOLOv4,and the detection speed is increased by 45.4%,the application prospect of this method is indicated in real-time detection of generator stator defects on an embedded platform.