首页|基于卷积神经网络和激光超声的表面缺陷检测

基于卷积神经网络和激光超声的表面缺陷检测

扫码查看
针对激光超声检测中表面缺陷深度、角度定量识别的问题,提出一种基于深层卷积神经网络(CNN)和支持向量机集成学习的近表面缺陷检测新方法.首先,通过建立有限元模型模拟不同缺陷深度、角度的激光超声信号,然后利用小波变换对信号进行时频分析,获取带有时域和频域特征的时频谱图,最后将时频谱图输入深层CNN和支持向量机模型中进行缺陷深度、角度的预测.结果表明,所提模型能够对缺陷深度、角度进行高精度预测,回归系数保持在0.98以上,真实值与预测值的平均相对误差保持在13%以下,预测精度不仅优于集成学习中的单个网络模型,还优于常见的CNN神经网络模型和VRD集成神经网络模型.
Near-Surface Defect Detection Using Convolutional Neural Network and Laser Ultrasound Testing
A novel method for near-surface defect detection combining deep convolutional neural network(CNN)and support vector machine ensemble learning was proposed to address the challenge of quantitative recognition of surface defect depth and angle in laser ultrasonic testing.First,a finite element model was developed to simulate laser ultrasonic signals with various defect depths and angles.Second,wavelet transform was applied to perform time-frequency analysis of the signals,generating spectra that capture both time-and frequency-domain features.Finally,the time-frequency spectra were input into deep CNN and support vector machine models to predict defect depths and angles.Results reveal that the proposed model achieves high-precision predictions of defect depth and angle,with regression coefficients exceeding 0.98 and an average relative error between the true and predicted values below 13%.Notably,this prediction accuracy surpasses that of individual network models in ensemble learning,and other widely used CNN models and VRD ensemble neural network models.

measurement and metrologylaser ultrasoundfinite element simulationwavelet transformneural network

郭明泽、张兴媛、金桢玥

展开 >

上海工程技术大学航空运输学院,上海 201620

测量与计量 激光超声 有限元仿真 小波变换 神经网络

2024

激光与光电子学进展
中国科学院上海光学精密机械研究所

激光与光电子学进展

CSTPCD北大核心
影响因子:1.153
ISSN:1006-4125
年,卷(期):2024.61(23)