基于卷积神经网络的三维沉孔特征识别及关键参数提取  

3D countersink hole feature recognition and key parameter extraction based on convolutional neural network

作　　者：沈大为[1,2] 向华 庄新村[1,2] 赵震[1,2] Shen Dawei;Xiang Hua;Zhuang Xincun;Zhao Zhen(Institute of Forming Technology&Equipment,Shanghai Jiao Tong University,Shanghai 200030,China;National Engineering Research Center of Die&Mold CAD,Shanghai Jiao Tong University,Shanghai 200030,China)

机构地区：[1]上海交通大学塑性成形技术与装备研究院,上海200030 [2]上海交通大学模具CAD国家工程研究中心,上海200030

出　　处：《锻压技术》2022年第11期78-86,共9页Forging & Stamping Technology

基　　金：国家自然科学基金资助项目(51875351)。

摘　　要：精冲零件的工艺特征识别和关键参数提取是实现精冲工艺设计智能化的关键。针对典型精冲工艺特征——沉孔,构建了一个以三维CAD模型为输入的特征识别和参数提取模型。利用改进的自适应体素化算法,将基于参数驱动批量生成的沉孔CAD模型转化为体素化模型,建立模型样本数据集;采用两步法,分别以工艺特征体素化模型为输入建立基于三维卷积神经网络的沉孔特征识别模型和以沉孔中心截面图像为输入建立基于二维卷积神经网络的参数提取模型,依次实现了3类主要沉孔特征的分类识别和参数提取。经过验证和评估,所建模型对于沉孔特征类型识别与关键参数提取均有较高的准确率,可以为精冲工艺的智能化工艺设计提供有力支撑。Process feature recognition and key parameter extraction of fine blanking parts are the key point to realize intelligent process design for fine blanking. Therefore, for the typical fine blanking feature of countersink hole, a model for feature recognition and parameter extraction was constructed with 3 D CAD model as input. Then, using the improved adaptive voxelization algorithm, the CAD model of countersink hole generated in batches based on parameter-driven was converted into a voxelized model, and a data set of model sample was established. Furthermore, the two-step method was used by using the voxelization model of process features as input to establish a countersink hole feature recognition model based on 3 D convolutional neural network and using the center cross-section image of countersink hole as input to establish a parameter extraction model based on 2 D convolutional neural network, respectively, and the classification recognition and parameter extraction for three main types of countersink hole features were realized in turn. The results show that after verification and evaluation, the established model has high accuracy for the recognition of countersink hole feature types and the extraction of key parameters, which can provide strong support for the intelligent process design of fine blanking process.

关 键 词：沉孔 体素化 卷积神经网络 特征识别 参数提取 

分 类 号：TG386[金属学及工艺—金属压力加工]