基于神经网络学习及残余压痕形貌获取金属塑性力学参数

Acquisition of Metal Plastic Parameters Based on Neural Network Learning and Residual Indentation Morphology

压入法相比于其他传统力学测试方法具有试样加工简单及可实现原位测试等优势,区别于已有的基于压入载荷-深度曲线获取材料力学性能参数的方法,本文提出了一种基于残余压痕形貌及神经网络学习反演金属塑性力学参数的新方法;采用Instron万能材料试验机开展了紫铜、镁合金及低碳钢的球形压入测试,并通过轮廓形态系统对压入测试后的残余压痕形貌进行特征扫描以作为后续研究的数据基础,分析所提取数据的特点并进行放大、取整、二进制化及高位补充等处理;基于Abaqus二次开发自动提取不同材料参数模拟下的残余压痕深度数据并用于神经网络学习,比较并选取激活函数、初始化神经网络参数的方法、神经网络参数更新方式、损失函数、寻找最优参数策略及神经网络结构,使神经网络学习达到较好的效果;结合实验所得残余压痕形貌特征数据与学习后的神经网络得到紫铜、镁合金及低碳钢的相关塑性参数,将通过Instron万能材料试验机拉伸测试表征所得紫铜、镁合金及低碳钢的相关塑性参数值作为对照标准,得到了神经网络学习结果的相对误差,验证了所提出基于神经网络学习及残余压痕形貌获取金属塑性力学参数方法的有效性;该方法可推广到其他金属/合金材料的力学性能表征及塑性参数获取研究中.

Compared to conventional mechanical testing methods,the indentation method offers the advantages of simple manufacturing of samples and in-situ testing.This study proposes an alternative to deriving material mechanical parameters solely from indentation load-depth curves.It introduces an effective method for deducing metal plastic mechanical parameters based on residual indentation morphology and neural network learning.An Instron universal material testing machine was used to conduct spherical indentation tests on Cu,Mg,and Fe,followed by scanning their residual indentation morphology through the contour morphology system.The extracted morphology features served as the basis for further analysis.Data processing techniques such as amplification,rounding,binarization,and high-order digit supplementation were applied to the acquired data.Through Abaqus software and numerical simulations,residual indentation depth data associated with various material parameters were automatically extracted for neural network learning.Selections of activation function,neural network parameter initialization and updating mode,loss function,parameter optimization strategy,and neural network structure were carefully conducted to ensure effective learning.The plastic mechanical parameters of Cu,Mg,and Fe were obtained based on the residual indentation morphology feature data from indentation tests and the neural networks after learning.Additionally,the related plastic mechanical parameters of Cu,Mg,and Fe were also acquired through conventional uniaxial tensile tests and characterization using the Instron machine.By comparing the neural network learning results with tensile test data,relative errors in plastic mechanical parameters were identified.The effectiveness of the proposed method in obtaining metal plastic mechanical parameters based on neural network learning and residual indentation morphology was validated.This method can be expanded for characterizing mechanical properties and acquiring plastic parameters of other metal/alloy materials.