Online Monitoring Machining Errors of Thin-Walled Workpiece: A Knowledge Embedded Sparse Bayesian Regression Approach
Online Monitoring Machining Errors of Thin-Walled Workpiece: A Knowledge Embedded Sparse Bayesian Regression Approach
复制标题
DOI：
10.1109/tmech.2019.2912195
复制
发表时间：
2019-04
期刊：
影响因子：
--
通讯作者：
Le Cao;Xiaoming Zhang;Tao Huang;H. Ding
中科院分区：
文献类型：
--
作者：
Le Cao;Xiaoming Zhang;Tao Huang;H. Ding
切削力引起的刀具和工件的变形是薄壁工件加工误差的主要原因。监测这种力引起的误差对于真实的实时控制和补偿与变形相关的加工故障起着极其重要的作用。然而，现有的解析预测方法计算量大、计算复杂，难以满足在线预测加工误差的要求。因此，数据驱动的回归方法被引入到在线预测的加工误差。挑战在于：首先，需要通过有限的测量点来构造加工误差的空间连续分布;其次，回归模型必须具有较高的泛化性能，以适应不同的切削参数;第三，模型的复杂性应受到限制，以提高实时性。为了解决这些问题，提出了一种嵌入知识的回归模型，以建立加工误差与切削参数、切削位置以及在线测量的切削力之间的关系。模型中考虑了加工误差产生的物理机制，提高了模型的泛化精度。为了在有限样本的情况下学习权值变量，提高预测效率，在模型权值上引入高斯先验分布以减少模型冗余。结果表明，所提出的模型预测的加工误差符合测量比纯数据依赖回归预测。
Deflection of the tool and workpiece caused by cutting forces usually leads to machining errors of thin-walled workpieces. Monitoring this kind of force-induced errors plays an extremely important role in controlling and compensating the deflection-related machining failures in real time. However, accompanied by time consuming and complicated computations, now available analytical prediction methods cannot satisfy the requirements of online machining errors prediction. Hence, data-driven regression methods are introduced to online predict the machining errors. The challenges lie in that: first, the spatial continuous distribution of machining errors needs to be constructed via limited measured points; second, the regression model must have high generalization performance to adapt varied cutting parameters; and third, the model complexity should be restrained to improve the real-time performance. To tackling these challenges, a knowledge embedded regression is presented to model the relationship between machining error and cutting parameters, cutting location, as well as online measured cutting forces. The physical mechanism about machining errors is integrated into the model for improving the generalization accuracy. A Gaussian prior distribution over the model weights is introduced to reduce the model redundancy, for the sake of learning the weight variables with limited samples and increasing the prediction efficiency. Results have indicated that the predicted machining errors by the proposed model accords with the measurement better than those predicted by a purely data-dependent regression.