基于迁移学习的飞机线束模板图纸识别模型研究  

Research on Recognition Model of Aircraft Wiring Harness Template Drawings Based on Transfer Learning

作　　者：周文举[1] 朱一丁 杨庆华 朱喜 王长法 ZHOU Wenju;ZHU Yiding;YANG Qinghua;ZHU Xi;WANG Changfa(School of Mechatronic Engineering and Automation,Shanghai University,Shanghai 200444,China;Shanghai Aircraft Design and Research Institute,COMAC,Shanghai 201210,China)

机构地区：[1]上海大学机电工程与自动化学院,上海200444 [2]中国商飞上海飞机设计研究院,上海201210

出　　处：《信阳师范学院学报（自然科学版）》2024年第2期240-245,共6页Journal of Xinyang Normal University(Natural Science Edition)

基　　金：国家自然科学基金项目(62005155);上海市科技攻关“揭榜挂帅”项目(YDZX20223100004008)。

摘　　要：针对航空电气线路互联系统(EWIS)线束模板图纸在更改时的识别与分类问题,设计了一种基于迁移学习的飞机线束模板图纸内容分类方法。由现有的线束模板图纸建立线束模板图纸元素数据集,选定VGG-16网络作为基础网络进行改进,优化模型结构与参数。为了解决数据集规模较小并提高模型的识别精度及训练效率的问题,将大规模数据集训练的预训练模型迁移至线束模板图纸元素数据集进行训练,并比较不同迁移学习方式及不同网络结构对模型性能的影响。试验结果表明,微调迁移学习模型具有较高的分类准确度,达到了98.89%。与其他网络相比,该微调迁移学习模型具有更短的训练时长以及更高的识别准确率,具有较好的应用前景。A transfer learning-based classification method for aircraft harness template drawing content was designed for the recognition and classification of Electrical Wiring Interconnection Systems(EWIS)harness template drawings during change review.The harness template drawing element dataset was established from the existing harness template drawings,the VGG-16 network was selected as the basic network to optimize the model structure and parameters.To solve the problem of small dataset size and to improve the recognition accuracy and training efficiency of the model,the pre-trained model trained in the large-scale dataset was migrated to the harness template drawing element dataset for training,and the effects of different migration learning methods and different network structures on the model performance were compared.The experimental results showed that the fine-tuning transfer learning model had a high classification accuracy of 98.89%.Moreover,compared with other networks,the fine-tuning transfer learning model had a shorter training time and higher recognition accuracy,which has better application prospects.

关 键 词：航空电气线路互联系统(EWIS) 线束模板图纸 卷积神经网络 VGG-16网络 迁移学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] V242[自动化与计算机技术—计算机科学与技术]