基于深度神经网络的7065铝合金厚板应力检测模型  

Stress detection model of 7065 aluminum alloy thick plate based on deep neural network

作　　者：杨小平[1,3] 武修瑞 郑许[4,5] 任月路 朱玉涛[4,5] 何克准 卢祥丰[4,5] 莫红楼 YANG Xiaoping;WU Xiurui;ZHENG Xu;REN Yuelu;ZHU Yutao;HE Kezhun;LU Xiangfeng;MO Honglou(College of Physics and Electronic Information Engineering,Guilin University of Technology,Guilin 541004,China;College of Computer Science and Technology,Guilin University of Technology,Guilin 541004,China;Guangxi Key Laboratory of Embedded Technology and Intelligent System,Guilin University of Technology,Guilin 541004,China;ALG Aluminum Inc,Nanning 530031,China;Guangxi Key Laboratory of Materials and Processes of Aluminum Alloys,Nanning 530031,China)

机构地区：[1]桂林理工大学物理与电子信息工程学院,广西桂林541004 [2]桂林理工大学计算机科学与工程学院,广西桂林541004 [3]桂林理工大学广西嵌入式技术与智能系统重点实验室,广西桂林541004 [4]广西南南铝加工有限公司,广西南宁530031 [5]广西铝合金材料与加工重点实验室,广西南宁530031

出　　处：《铁道科学与工程学报》2024年第9期3787-3796,共10页Journal of Railway Science and Engineering

基　　金：广西科技重大专项项目(桂科AA23023027);广西重点研发计划项目(桂科AB23026149)。

摘　　要：针对工业生产中传统超声应力检测法对铝合金厚板在不同拉伸率和不同温度条件下存在的测量误差的问题,以7065铝合金厚板为实验对象,提出一种在不同拉伸率和不同温度条件下的基于树突神经网络的应力预测模型与传统超声检测法融合的应力检测模型,然后使用改进的GSA-GRNN对该应力检测模型进行温度补偿。以南南铝公司生产的7065铝合金厚板为研究对象,使用恒温槽为超声检测提供恒温环境,分别对不同拉伸率、不同温度下的7065铝合金厚板进行超声检测,将声时差、拉伸率作为输入参数,应力作为输出参数,创建一个基于树突神经网络的应力检测模型,然后将应力检测模型的输出作为输入,使用改进的GSA-GRNN建立温度补偿模型对应力检测模型进行温度补偿。研究结果表明:融合了传统超声声时差的检测模型均方根误差为0.84636,相关系数为0.99743,和其他神经网络模型对比,该模型拥有更好的精度;在对该模型进行温度补偿后,模型的应力均方根误差和相关系数分别可以达到0.78848和0.99844,模型的精度得到了进一步的提升。证明基于数据驱动的神经网络融合传统超声检测可以有效降低检测误差,同时省去传统检测方法人工计算应力的时间,提高了检测效率。研究结果可以为基于数据驱动的应力检测模型提供进一步的优化参考。Aiming at the problem of measuring error of the traditional ultrasonic stress detection method on the aluminum alloy thick plate under different tensile ratio and different temperature conditions in industrial production,a stress detection model based on the dendritic neural network stress prediction model and the traditional ultrasonic detection method under different tensile rates and different temperature conditions was proposed.The improved GSA-GRNN was used to compensate the temperature of the stress detection model.In this study,7065 aluminum alloy thick plate produced by Nannan Aluminum Company was taken as the research object,and a constant temperature tank was used to provide a constant temperature environment for ultrasonic testing.Ultrasonic testing was carried out on 7065 aluminum alloy thick plate under different tensile rates and different temperatures,and the sound time difference and tensile rate were taken as input parameters and the stress as output parameters.A stress detection model based on dendritic neural network was created.Then,the output of the stress detection model was taken as input,and the improved GSA-GRNN was used to build a temperature compensation model to compensate the stress detection model.The results are shown as follows.The root-mean-square error and correlation coefficient of the detection model integrating the traditional ultrasonic sound time difference are 0.84636 and 0.99743.Compared with other neural network models,this model can have better accuracy.After temperature compensation,the stress root-mean-square error and correlation coefficient of the model can reach 0.78848 and 0.99844,respectively.The accuracy of the model has been further improved,which proves that the data-driven neural network integrated with traditional ultrasonic detection can effectively reduce the detection error,save the time of manual stress calculation of traditional detection methods,and improve the detection efficiency.The research results can provide further optimization reference for stre

关 键 词：应力检测 树突神经网络 粒子群算法 万有引力搜索算法 声时差 

分 类 号：TH142[一般工业技术—材料科学与工程] TP391.9[机械工程—机械制造及自动化] U214[自动化与计算机技术—计算机应用技术]