基于机器学习的管材数控弯曲质量预测

Tube numerical controlled bending quality prediction based on machine learning

在管材数控(NC)弯曲过程中,可能出现起皱、过度减薄的质量缺陷,同时会不可避免地发生回弹,都将严重影响成形质量。为了对数控弯曲成形质量进行预测,提出了使用有限元模拟与机器学习相结合的方法,并建立了快速的成形质量预测方法。首先,建立了有效的管材数控弯曲的参数化有限元模型,在工艺参数取值范围中随机选择进行大量的模拟实验作为样本,完成学习数据的挖掘。随后,基于径向基函数(RBF)神经网络建立壁厚减薄与回弹程度的预测模型并使用支持向量机(SVM)建立管材起皱的预测模型。最后,使用模型对新的实例进行预测,并利用模拟与数控弯曲实验对预测模型进行验证。该方法可以对大直径薄壁管材数控弯曲质量进行有效的预测,提高弯曲管件零件设计效率。

Wrinkle and over thinning, as well as inevitable spring-back, may occur along with the tube numerical controlled （NC） bending process, which have strong impacts on forming quality. As the bending processing is a complex non-linear system, it is hard to compute the result theoretically. Besides, finite element simulation is a time-consuming method for industry. To predict the forming quality of the NC bending, a rapid method based on the machine learning method and finite element modeling is raised. To apply the meth- od, the first step is to build the finite element model of tube bending and make simulations whose process pa- rameters are selected randomly as samples. After extracting experimental data, a radius basis function （RBF） neural network and a support vector machine （SVM） are built to predict thinning, spring-back and wrinkle separately. New instances are taken to verify the prediction method. The results show that the machine learning method can reliably predict the large diameter thin-walled tube NC bending quality and improve the efficiency of part forming process design.