摘要
针对齿轮表面缺陷识别过程中特征信息提取不充分和识别精度较低等问题,提出一种融合多尺度-注意力机制和迁移学习的齿轮表面缺陷识别方法。新方法的创新之处包括:在AlexNet网络中引进Inception模块增加模型宽度,提取多尺度信息,从而提升模型对于不同尺度缺陷特征的感知能力;在多尺度编码操作前,采用卷积注意力机制模块在通道维度和空间维度增强缺陷特征信息,进一步提升模型的特征表达能力;将在通用数据集上预训练的AlexNet模型中部分参数信息迁移至网络模型,利用权重微调策略实现针对齿轮表面缺陷识别任务的个性化网络参数优化;最后,将所提方法在齿轮表面缺陷测试集上进行测试。实验结果表明,该方法在齿轮表面缺陷测试集上的精确率为99.33%,召回率为99.33%,分别提升了9.13%和13.23%,满足实际工业生产中齿轮表面缺陷识别的需求。
To address the issues like insufficient feature extraction and low accuracy in the recognition task of gear surface defects,a new gear surface defects recognition method based on multi-scale attention mechanism and transfer learning was proposed.The main novelties of the new method are as follaws:the Inception module is included in the AlexNet network architecture to increase the width of the model and extract multi-scale features,which can improve the sensory capability of the model;the convolutional attention mechanism is adopted to strengthen the defect features in the channel dimension and space dimension,which can further enhance the feature representation ability;partial parameters in the new model is transferred from an AlexNet pre-trained on a general dataset and fine-tuned for characteristic optimization of the gear surface defect recognition task.Finally,the proposed method was tested on the test set of gear surface defects.The experimental result indicates that the precision of the method on the test set is 99.33%and the recall rate is 99.33%,improving by 9.13%and 13.23%respectively.It satisfies demands of gear surface defect recognition in the actual industrial production.
作者
高齐
尹明锋
吴祥
符诗语
贝绍轶
GAO Qi;YIN Mingfeng;WU Xiang;FU Shiyu;BEI Shaoyi(College of Mechanical Engineering,Jiangsu University of Technology,Changzhou 213001,China;School of Automobile and Traffic Engineering,Jiangsu University of Technology,Changzhou 213001,China;School of Automation,Nanjing University of Science and Technology,Nanjing 210094,China)
出处
《现代制造工程》
CSCD
北大核心
2023年第8期7-14,98,共9页
Modern Manufacturing Engineering
基金
国家自然科学基金项目(62103192)
江苏省高等学校自然科学研究面上项目(20KJB520015)
常州市应用基础研究计划项目(中补助)(CJ20200039)。
