转向架构架表面缺陷的磁粉探伤检测算法研究

Research on Magnetic Particle Inspection Testing Algorithm for Surface Defects of Bogie Frame

中文关键词:  转向架构架  磁粉探伤  缺陷检测  YOLO-CET  注意力机制

英文关键词:bogie frame  magnetic particle inspection  defect detection  YOLO-CET  attention mechanism

中文摘要:

      针对转向架构架磁粉探伤缺陷识别环节人工目测效率低的现状，提出一种基于YOLO-CET（You Only Look Once based on CoTNet-Efficient-Transformer blocks）的探伤图像缺陷自动识别算法，实现对构架表面真伪缺陷的智能识别。以YOLOv5(You Only Look Once version 5)为基础模型，在骨干特征提取网络引入轻量化CoTNet(Contextual Transformer Networks)网络层，实现缺陷特征的多尺度融合与提取。加入高效通道注意力机制，在不增加网络计算量的同时提高模型的鲁棒性和泛化性。增加一个小尺寸缺陷检测头用于减轻不同尺寸特征带来的尺度方差影响，同时引入视觉自注意力模块，增强小目标缺陷的抓取识别能力。利用自建的构架表面缺陷探伤数据集进行测试，结果表明，与YOLOv5相比，所提出的YOLO-CET使检测平均精度提升33.8%，F1-Score提升0.26，浮点运算量仅增加1.5 B，该模型可实现缺陷的自动检测，有效解决背景误判、细小缺陷漏检等问题。

英文摘要:

      A YOLO-CET（You Only Look Once based on CoTNet-Efficient-Transformer blocks） based automatic defect recognition algorithm for magnetic particle inspection of bogie frames is proposed to address the low efficiency of manual visual inspection in defect recognition.This algorithm achieves intelligent recognition of true and false defects on the surface of the frame.Based on the YOLOv5(You Only Look Once version 5)model,a lightweight CoTNet(Contextual Transformer Networks) layer is introduced into the backbone feature extraction network to achieve multi-scale fusion and extraction of defect features.By incorporating an efficient channel attention mechanism,the robustness and generalization of the model can be improved without increasing the computational complexity of the network.A small-sized defect detection head is added to alleviate the scale variance caused by features of different sizes,and a visual self attention module is introduced to enhance the ability to capture and recognize small target defects.A self built dataset for frame surface defect inspection is used for testing,the results show that compared with YOLOv5,the proposed YOLO-CET improves the average detection accuracy by 33.8%,the F1-Score by 0.26 and the floating point operation by only 1.5 B.This model can achieve automatic defect detection and effectively solve problems such as background misjudgment and missed detection of small defects.