华中科技大学学报（自然科学版）2025，Vol.53Issue(6)：19-27,9.DOI:10.13245/j.hust.250604
基于跨图匹配推理的齿轮端面缺陷定位检测
Location and detection of gear end face defects based on cross-graph matching reasoning
摘要
Abstract
Aiming at the problem of low-precision gear end face defect location detection in the actual production line environment,a cross-graph matching reasoning method was proposed.First,to improve the adaptability of the location model to different environments,the YOLO(you only look once)model architecture under the adversarial mechanism was analyzed,and by analyzing the relationships between image-level features,a step-by-step feature compression structure was constructed to enhance the discrimination ability of the image-level domain classifier.Then,an adaptive gradient reversal layer was established between the domain classifier and the feature layer to deeply mine the information of hard samples and enhance the generalization ability of the model.Finally,the distribution characteristics of gear instance-level features were explored.Based on the graph convolutional network,a feature map structure was constructed,and during the location process,a graph-matching criterion based on the adversarial mechanism was established to guide the model to focus on domain-invariant features and realize knowledge transfer in different environments.Experimental results show that,compared with other methods,the proposed method achieves higher precision in gear end face defect localization and detection.关键词
齿轮缺陷检测/YOLO定位模型/对抗学习/图卷积网络/跨图匹配推理Key words
gear defect detection/YOLO localization model/adversarial learning/graph convolutional network/cross-graph matching reasoning分类
机械工程引用本文复制引用
卢义,赵新维,薛志钢,顾杰斐,宿磊,李可..基于跨图匹配推理的齿轮端面缺陷定位检测[J].华中科技大学学报（自然科学版）,2025,53(6):19-27,9.基金项目
国家重点研发计划资助项目(2023YFB4404203) （2023YFB4404203）
国家自然科学基金资助项目(52375099,U23B2044). （52375099,U23B2044）
