基于回归方法和神经网络的电弧增材制造单道成形形貌预测

目的 针对电弧增材制造技术实际应用中工艺参数选取困难和成形结果难预测的问题，确定高效、准确的电弧增材制造单道成形形貌预测的数学方法，以快速、方便地选取丝材电弧增材制造工艺参数并指导成形质量控制。方法 在单道单层丝材电弧增材制造实验的基础上，采用多种回归方法和神经网络方法分别建立焊接电流、电压和焊枪移动速度等多个工艺参数与增材层宽度、增材层高度及熔池深度等成形形貌参数之间的数学关系模型。结果 电弧增材制造单道成形形貌与焊接电流、电压和焊枪移动速度显著相关，且各参数间存在非线性交互作用;采用多元线性回归法可较准确地预测单道增材层宽度，但对于增材层高度和熔深的预测效果较差;神经网络可良好地处理各工艺参数间复杂的非线性关系，其对增材层宽度、增材层高度和熔深的预测平均误差率分别为4.17%、6.60%和7.01%，显著优于多元线性回归法。结论 采用神经网络法可以准确预测电弧增材制造单道成形的形貌参数，进而指导增材制造工艺参数的选取和成形质量的控制。

Prediction of Single-pass Arc Additive Manufacturing Forming Morphology Based on Regression and Neural Network

The work aims to determine an efficient and accurate mathematical method for predicting the forming morphology of single pass in arc additive manufacturing to solve the difficulties in selecting process parameters and predicting forming results in practical application of arc additive manufacturing technology, so as to select the process parameters of wire arc additive manufacturing quickly and conveniently and guide the forming quality control. A variety of regression methods and neural network methods were used to establish the mathematical relationship model between the multiple process parameters (welding current, voltage, welding torch moving speed, et al) and forming morphology parameters (width of additive layer, height of additive layer and depth of weld penetration) based on the single-pass single-layer wire arc additive manufacturing experiment. The results showed that the single-pass forming morphology of arc additive manufacturing was significantly related to welding current, voltage and moving speed of welding torch, and nonlinear interaction existed between the parameters. The multiple linear regression method could accurately predict the width of single additive layer, but the prediction effect of additive layer height and weld penetration depth was poor. The neural network could handle the complex nonlinear relationship among the process parameters well, its average prediction error rates on the width, height and penetration of the additive layer were respectively 4.17%, 6.60% and 7.01%, which were significantly lower than those of the multiple linear regression method. The neural network method can accurately predict the morphology parameters of arc additive manufacturing single-pass forming, and then guide the selection of additive manufacturing process parameters and the control of forming quality.