仿真数据驱动的起重机钢丝绳断丝定量识别方法  

Quantitative identification method of broken wires of crane wire ropes driven by simulation data

作　　者：王晓昆 井陆阳 白晓瑞 战卫侠 刘云成 王彦松 WANG Xiaokun;JING Luyang;BAI Xiaorui;ZHAN Weixia;LIU Yuncheng;WANG Yansong(Mechanical and Automotive Engineering College,Qingdao Technological University,Qingdao 266520,China;College of Weaponry Engineering,Naval University of Engineering,Wuhan 430032,China;MESNAC Co.,Ltd.,Qingdao 266000,China)

机构地区：[1]青岛理工大学机械与汽车工程学院,山东青岛266520 [2]海军工程大学兵器工程学院,湖北武汉430032 [3]软控股份有限公司,山东青岛266000

出　　处：《机电工程》2024年第1期43-51,共9页Journal of Mechanical & Electrical Engineering

基　　金：山东省自然科学基金资助项目(ZR2020QE158,ZR2021ME026);山东省科技型中小企业创新能力提升工程项目(2021TSGC1045)。

摘　　要：针对工程应用中起重机钢丝绳损伤检测由于故障样本不足导致检测效果较差的问题,提出了一种仿真数据驱动的、基于卷积神经网络(CNN)模型的钢丝绳断丝定量识别方法。首先,使用有限元软件COMSOL对直径24 mm的6×37结构钢丝绳不同损伤类型进行了建模与仿真,并将提离值为5 mm处仿真漏磁场减去背景磁场,得到了仿真损伤信号;其次,搭建了钢丝绳漏磁检测试验台,采集了与仿真参数一致的钢丝绳断丝实测信号,并进行了小波去噪处理,解决了实测信号中由于存在噪声干扰导致与仿真信号不匹配的问题;最后,建立了卷积神经网络模型,采用仿真数据辅助模型训练,并使用早停法抑制了模型对训练集中仿真数据的过拟合问题。以仿真数据作为训练集和仿真数据辅助小样本实测数据作为训练集,进行了实验验证。研究结果表明:仿真数据驱动、小波去噪能够较大幅度地提高断丝识别率;早停法能够较好地抑制实验中的过拟合问题;该方法在两种实验形式下的准确率分别达到了84.5%和97.5%,证明该方法能够有效识别钢丝绳损伤,具有一定的理论研究价值和工程应用前景。To solve the problem that damage detection of crane wire ropes suffers from insufficient fault samples in engineering,a quantitative identification method of broken wire ropes based on convolution neural network(CNN)model and driven by simulation data was proposed.Firstly,the finite element software COMSOL was used to model and simulate the different damage types of 6×37 structural wire ropes with a diameter of 24 mm,and the simulated leakage magnetic field was subtracted from the background magnetic field at the lift-off value of 5 mm to obtain the simulated damage signals.Secondly,the magnetic flux leakage test bench of wire ropes was set up to collect the measured signals of broken wire ropes consistent with the simulation parameters,and the wavelet denoising was used to solve the problem that the measured signals did not match the simulated signals due to noise interference.Finally,the convolution neural network model was established and the simulated data were used to assist the model training,and the early stop method was introduced to prevent the overfitting of the model to the simulated data.The experimental verification was conducted using simulation data as the training set and simulation data assisted with small sample measured data as the training set.The research results show that the recognition rate of broken wires can be greatly improved by the simulation data-driven and wavelet denoising,and the early stop method can effectively suppress overfitting problems in experiments.The accuracy of the method in two experimental forms is 84.5%and 97.5%respectively,which proves that the method can effectively identify the damage of wire ropes,and has certain theoretical research value and engineering application prospect.

关 键 词：起重机械 电磁检测 模式识别方法 卷积神经网络 小波变换 小样本数据 COMSOL 

分 类 号：TH21[机械工程—机械制造及自动化] TP183[自动化与计算机技术—控制理论与控制工程]