基于神经网络和仿真数据的粗糙度分类检测

Roughness Classification and Detection Based on Neural Network and Simulation Data

针对卷积神经网络对训练样本的需求量大,但单纯通过实测实验建立的粗糙度样本量有限的问题,提出了基于神经网络和仿真数据的粗糙度分类检测方法。首先,基于Z-Map算法生成铣削表面仿真形貌,获取粗糙度与实际粗糙度相近的仿真图像,提取铣削仿真图像和实际图像的纹理,采用深度迁移学习Deep CORAL模型对仿真和实际的纹理进行匹配;其次,在原有6类粗糙度数据集的基础上加入铣削表面仿真图像,扩充数据量,采用卷积神经网络Xception模型,对6类粗糙度进行分类识别。结果表明,与没有加入仿真图像的模型相比,准确率从86.48%提升到92.79%。加入的铣削仿真图片扩充了数据量,帮助模型提取到与粗糙度相关的更多有用的特征,使得分类检测结果提高。

In view of the large demand of convolutional neural network for training samples and the limited number of roughness samples simply established through actual measurement experiments,a method for milling surface roughness classification and prediction using neural network combined with simulation data was proposed.Firstly,the simulation topography of milling surface was generated based on Z-Map algorithm,and the simulation image of roughness close to actual roughness was obtained.The texture of the milled simulation and real images were extracted and matched by deep transfer learning Deep CORAL model.In addition,simulation images of milling surface were added to the original six roughness data sets to expand the amount of data.Compared with the model without simulation images,the accuracy of the six roughness classification was improved from 86.48%to 92.79%by using Xception model of convolutional neural network.The added milling simulation images expand the data amount,help the model to extract more useful features related to roughness,and improve the classification detection results.