基于深度学习和S试件的五轴机床误差溯源方法研究与实现

发布时间：2018-12-08 12:20

【摘要】：今年,我国大飞机C919完成首航,航母舰队投入使用,国人为之自豪。五轴机床在此发挥了重要作用。五轴机床在切削自由曲面时具备优异的性能,广泛应用于船舶、航空航天、国防军工等领域。五轴机床在三轴机床基础上增加了两个转动轴,提升加工性能的同时极大增加了机构运动的复杂性,目前国内外对于五轴机床的加工性能测试缺乏相应的评定标准。成都飞机工业公司根据国外检测试件缺陷和实际生产经验设计出S试件。通过五轴机床切削S试件,能综合反映机床的加工性能和动态特性。但目前建立S试件与五轴机床误差项间的映射函数较为困难,基于S试件的机床误差溯源理论仍然有待完善。深度学习是一种机器学习算法,能够直接处理图像等高维结构数据,自动建立准确的映射函数,在多个相关领域的取得成功。本文在现有误差溯源方法基础上,参考深度学习在相关任务的应用办法,提出一种基于S试件和深度学习的五轴机床误差溯源方法,对于完善S试件的机床误差溯源理论并提高五轴机床的加工性能具有重要意义。具体研究内容如下:1、设计实际刀具位姿切削S试件的仿真算法,建立五轴机床单项误差与S试件轮廓度误差的正向映射函数。包括基于多体理论建立五轴机床空间误差模型、通过Matlab辅助分析各单项误差与S件轮廓度误差的对应关系。2、设计用于拟合S试件与误差项间映射函数的深度学习卷积神经网络结构,特殊改进包括将S试件三维点云数据通过拉伸和投影转化为三张不同维度的误差图,为此设计三通道的结构;误差数据的分布不稳定,精度高,采用批量正则化操作调控数据分布;实现剪枝策略对网络结构进行优化。3、实现基于深度学习和S试件的五轴机床误差溯源实验方案,包括规划数据集训练策略、规格标准;基于ipython notebook建立可在线编程、可视化的客户端测试框架;搭建维护实验室深度学习服务器与远端登陆等服务;最后基于Caffe平台进行实验,验证了整体方案的可行性和正确性。
[Abstract]:This year, China's large aircraft C919 completed its first flight, aircraft carrier fleet put into use, Chinese pride. Five-axis machine tools play an important role here. Five-axis machine tool has excellent performance in cutting free surface. It is widely used in ship, aerospace, national defense industry and so on. The five-axis machine tool adds two rotating shafts on the basis of three-axis machine tool, which improves the machining performance and greatly increases the complexity of mechanism movement. At present, there is no corresponding evaluation standard for the testing of machining performance of five-axis machine tool at home and abroad. Chengdu aircraft Industry Company designs S specimen according to the foreign test sample defect and actual production experience. The machining performance and dynamic characteristics of the machine tool can be synthetically reflected by cutting S specimen with five axis machine tool. However, it is difficult to establish the mapping function between S-specimen and five-axis machine tool error term at present, and the theory of machine tool error traceability based on S-specimen still needs to be perfected. Depth learning is a machine learning algorithm, which can directly process high-dimensional structural data such as images, automatically establish accurate mapping functions, and succeed in many related fields. On the basis of existing error tracing methods and referring to the application of depth learning in related tasks, this paper presents a method of error traceability for five-axis machine tools based on S specimen and depth learning. It is of great significance to improve the error tracing theory of S-specimen and improve the machining performance of five-axis machine tool. The main contents of this paper are as follows: 1. The simulation algorithm of cutting S specimen is designed, and the forward mapping function between the single error of five axis machine tool and the contour error of S specimen is established. The spatial error model of five-axis machine tool is established based on multi-body theory, and the corresponding relationship between the individual error and the profile error of S part is analyzed by Matlab. 2. Deep-learning convolution neural network structure is designed to fit the mapping function between S specimen and error item. Special improvements include transforming S sample 3D point cloud data into three different dimension error maps by stretching and projecting. The three-channel structure is designed for this purpose. The distribution of error data is unstable and accurate, so batch regularization operation is used to control the data distribution. The pruning strategy is realized to optimize the network structure. 3. To realize the experimental scheme of error traceability of five-axis machine tool based on depth learning and S sample, including planning data set training strategy and specification standard; Based on ipython notebook, the test framework of online programming and visual client is established; the service of maintenance laboratory depth learning server and remote landing is built; finally, the feasibility and correctness of the whole scheme are verified based on the Caffe platform.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TG659

【参考文献】