摘要
目的为了快速识别机械加工表面的三维高度参数偏度S_(sk)、峰度S_(ku),基于谱表示法(Spectral Representation Method,SRM)模型构建了粗糙表面数据集,并利用卷积神经网络(Convolutional Neural Network,CNN)对粗糙表面的偏度、峰度进行快速识别。方法首先采用KEYENCE VK-X260K激光共聚焦显微系统测试了立铣、平铣、平磨表面形貌数据,提取真实加工表面合理的峰度、偏度范围以确定神经网络输入的偏度、峰度范围,并在此范围基础上使用谱表示法构建了一个包含不同偏度、峰度的三维粗糙表面数据集。然后通过单因素实验法分析网络参数(网络深度,滤波器大小)对偏度、峰度识别精度的影响,以寻求最优的神经网络参数组合,最后通过与传统统计特征计算法,即基于谱表示法模型生成粗糙表面后计算其表面高度分布的三阶中心矩和四阶中心矩进行对比,验证卷积神经网络法识别三维参数偏度、峰度的有效性。结果实验结果表明,基于卷积神经网络方法计算的偏度、峰度最优绝对百分比误差达到8.8%和1.5%,其平均百分比误差分别可以控制在12.9%和3.7%以内。结论所建立的非高斯粗糙表面能准确描述形貌特征,能正确、清晰地反映三维高度参数,且卷积神经网络对粗糙表面参数的识别具有优势性。研究结果为粗糙表面高度特征参数的识别提供了新思路。
The rapid and accurate identification of three-dimensional(3D)height parameters,specifically skewness S_(sk)and kurtosis S_(sk).is critical for evaluating the functional performance of manufacturing.Traditional approaches for quantifying these parameters often rely on labor-intensive experimental measurements or computationally expensive numerical simulations,which hinder real-time quality control and process optimization.To address these limitations,this study presents an integrated framework combining the spectral representation method(SRM)with a convolutional neural network(CNN)to automate the identification of S_(sk)and S_(ku)for non-Gaussian rough surfaces.The proposed methodology not only accelerates parameter estimation but also bridges the gap between stochastic surface modeling and data-driven machine learning.The research commenced with the acquisition of high-fidelity surface topography data from three common machining processes:vertical milling,face milling,and flat grinding.A KEYENCE VK-X260K laser confocal microscope was employed to capture 3D surface profiles at a resolution of 0.1μm,ensuring precise measurement of micro-scale asperities.Statistical analysis of the experimental data revealed that the realistic ranged for skewness and kurtosis in industrial contexts span S_(sk):−0.7 to 0.7 and S_(ku):1.5 to 5.0,respectively.These ranges were subsequently used to guide the generation of a synthetic dataset via the SRM,a stochastic modeling technique that reconstructed non-Gaussian surfaces by modulating spectral density functions and higher-order statistical moments.The dataset comprised over 10000 synthetic 3D surfaces,with each annotated with ground-truth S_(sk)and S_(ku)values,thereby providing a robust foundation for training and validating deep learning models.A dedicated CNN architecture was engineered to directly process 3D surface height matrices and predict skewness and kurtosis.The network design incorporated multi-scale convolutional layers(3×3,5×5,7×7,and 9×9 filters)to capture both local roughness features and global texture patterns,while batch normalization reduced the problem of gradient disappearance and accelerated the learning process of the model.In order to optimize the performance,a single factor experimental design was used to systematically evaluate the influence of network depth and filter size on the prediction accuracy.The convolutional neural network model was trained with an Adam optimizer,and the learning rate was set to 0.001.The results show that with the increase of network depth,the network performance gradually increases to the peak,and then decreases.However,the increase of filters in the network degrades the performance of the model.Only when the depth is 50 layers convolution and the filter size is 3×3,the network has the best performance.Compared with the traditional statistical feature calculation method,which calculates the third-order and fourth-order central moments of surface height distribution after generating a rough surface based on the spectral representation model.Experimental results show that for the skewness(S_(sk))and kurtosis(S_(ku))calculated based on the Convolutional Neural Network(CNN),the optimal absolute percentage errors(APEs)reach 8.8%and 1.5%respectively,and their mean absolute percentage error(MAPEs)can be controlled within 12.9%and 3.7%respectively.Notably,the CNN reduces computational time by three orders of magnitude,enabling near-instantaneous parameter estimation compared with the SRM simulations.This study not only advances the understanding of non-Gaussian surface generation and characterization,but also establishes a new paradigm for leveraging deep learning in tribology and precision manufacturing.The proposed methodology holds significant potential for real-time quality control,surface optimization,and functional performance prediction in industrial settings.
作者
王子杰
雷声
汪刘群
WANG Zijie;LEI Sheng;WANG Liuqun(School of Computer Science,South-Central Minzu University,Wuhan 430070,China)
出处
《表面技术》
北大核心
2025年第22期110-118,共9页
Surface Technology
基金
国家自然科学基金(52105135)
湖北省自然科学基金(2020CFB174)
南通市科技计划项目(JC2023008)。
关键词
偏度
峰度
卷积神经网络
接触表面
skewness
kurtosis
convolutional neural network
contact surface
