基于卷积神经网络的拧紧曲线异形识别方法研究
古洪亮，丁建明
（西南交通大学 牵引动力国家重点实验室，四川 成都 610031）
摘要：针对现有拧紧曲线异形识别方法依赖人工提取特征和专业知识的问题，提出了一种基于卷积神经网络的拧紧曲线异形识别方法。首先，针对拧紧曲线数据匮乏、数据长度不均的问题，通过随机裁剪来进行数据增强，通过回归决策树重构曲线实现曲线对齐；然后，基于传统CNN模型，建立了拧紧曲线异形识别模型；最后，通过研究各项超参数与模型识别结果的关系，给出了最终的参数组合方案，并且通过分析训练过程中各个阶段的混淆矩阵，展示了模型的学习过程，通过与传统机器学习方法SVM等对比，验证了本文所提出方法的有效性。实验结果表明：该方法能够有效地识别出螺栓连接拧紧曲线的异常数据，识别正确率可以达到99%，为基于深度学习的螺栓连接件智能诊断提供了方法指导。
关键词：卷积神经网络；螺栓连接；拧紧曲线；故障诊断
中图分类号：U270.6+7 文献标志码：A doi：10.3969/j.issn.1006-0316.2023.07.003
文章编号：1006-0316 (2023) 07-0019-07
Recognition Method of Abnormal Shape of Tightening Curve Based on Convolutional Neural Network
GU Hongliang，DING Jianming
( State Key Laboratory of Traction Power, Southwest Jiaotong University, Chengdu 610031, China )
Abstract：To address the problem that the existing methods largely depend on artificial feature extraction and professional knowledge, this paper proposes a method for identifying the abnormal shape of tightening curve based on convolutional neural network. Firstly, in view of the lack of data and uneven data length of tightening curve, the paper uses random clipping to enhance the data, and uses regression decision tree to reconstruct the curve to keep the same length. Secondly, based on the traditional CNN model, the identification model of tightening curve abnormity is established. Finally, by studying the influence of super parameters on diagnosis results, the final parameter combination scheme is given. By analyzing the confusion matrix in each stage of the training process, the model learning process is demonstrated. By comparing with the traditional machine learning method, such as SVM, the effectiveness of the method proposed is verified. The experimental results show that the method can effectively identify the abnormal data of the bolt connection tightening curve, and the recognition accuracy is close to 99%, which provides a guidance for the intelligent diagnosis method of bolt connection based on deep learning.
Key words：convolutional neural networks；bolted connections；process curve；fault diagnosis
———————————————
收稿日期：2022-12-27
基金项目：国家重点研发计划（2020YFA0710902）
作者简介：古洪亮（1998－），男，四川泸州人，硕士研究生，主要研究方向为紧固件缺陷智能检测，E-mail：1312331780@qq.com；丁建明（1981－），男，四川平昌人，博士，副研究员，主要研究方向为机电设备的智能控制与大数据可视化研究。
（西南交通大学 牵引动力国家重点实验室，四川 成都 610031）
摘要：针对现有拧紧曲线异形识别方法依赖人工提取特征和专业知识的问题，提出了一种基于卷积神经网络的拧紧曲线异形识别方法。首先，针对拧紧曲线数据匮乏、数据长度不均的问题，通过随机裁剪来进行数据增强，通过回归决策树重构曲线实现曲线对齐；然后，基于传统CNN模型，建立了拧紧曲线异形识别模型；最后，通过研究各项超参数与模型识别结果的关系，给出了最终的参数组合方案，并且通过分析训练过程中各个阶段的混淆矩阵，展示了模型的学习过程，通过与传统机器学习方法SVM等对比，验证了本文所提出方法的有效性。实验结果表明：该方法能够有效地识别出螺栓连接拧紧曲线的异常数据，识别正确率可以达到99%，为基于深度学习的螺栓连接件智能诊断提供了方法指导。
关键词：卷积神经网络；螺栓连接；拧紧曲线；故障诊断
中图分类号：U270.6+7 文献标志码：A doi：10.3969/j.issn.1006-0316.2023.07.003
文章编号：1006-0316 (2023) 07-0019-07
Recognition Method of Abnormal Shape of Tightening Curve Based on Convolutional Neural Network
GU Hongliang，DING Jianming
( State Key Laboratory of Traction Power, Southwest Jiaotong University, Chengdu 610031, China )
Abstract：To address the problem that the existing methods largely depend on artificial feature extraction and professional knowledge, this paper proposes a method for identifying the abnormal shape of tightening curve based on convolutional neural network. Firstly, in view of the lack of data and uneven data length of tightening curve, the paper uses random clipping to enhance the data, and uses regression decision tree to reconstruct the curve to keep the same length. Secondly, based on the traditional CNN model, the identification model of tightening curve abnormity is established. Finally, by studying the influence of super parameters on diagnosis results, the final parameter combination scheme is given. By analyzing the confusion matrix in each stage of the training process, the model learning process is demonstrated. By comparing with the traditional machine learning method, such as SVM, the effectiveness of the method proposed is verified. The experimental results show that the method can effectively identify the abnormal data of the bolt connection tightening curve, and the recognition accuracy is close to 99%, which provides a guidance for the intelligent diagnosis method of bolt connection based on deep learning.
Key words：convolutional neural networks；bolted connections；process curve；fault diagnosis
———————————————
收稿日期：2022-12-27
基金项目：国家重点研发计划（2020YFA0710902）
作者简介：古洪亮（1998－），男，四川泸州人，硕士研究生，主要研究方向为紧固件缺陷智能检测，E-mail：1312331780@qq.com；丁建明（1981－），男，四川平昌人，博士，副研究员，主要研究方向为机电设备的智能控制与大数据可视化研究。
设为首页 | 加入收藏 | 免责条款
《机械》杂志版权所有 Copyright©2008-2012 Jixiezazhi.com All Rights Reserved
电话：028-85925070 传真：028-85925073 E-mail：jixie@vip.163.com
地址：四川省成都锦江工业开发区墨香路48号 邮编：610063
Powered by PageAdmin CMS