Abstract: Surface roughness and tensile strength are important indicators to measure the quality of Fused Deposition Modeling (FDM) parts.However,FDM process parameters are numerous,and there is a non-linear relationship between FDM process parameters and FDM molded part quality,so it is difficult for traditional methods to predict these two indicators.Therefore,a quality prediction method for FDM 3D printed parts was proposed to improve the prediction accuracy and stability,which combined Bayesian Optimization and Hyperband (BOHB) algorithm and Back Propagation (BP) neural network.The process parameters were used as input to the model,such as layer thickness,number of scans and fill interval.The BOHB algorithm was used to optimize the hyperparameter of BP neural network for obtaining BOHB-BP.Then,the experimental data of surface roughness and tensile strength were obtained by central composite design,and the accuracy and stability of the model were verified according to the leave-one-out method on this data set.The predictions of the model BOHB-BP with the models GA-BP and BP were compared with each other.The above results proved that the proposed method had better prediction accuracy and stability on different datasets.

Key words: fused deposition modeling, quality prediction, Bayesian optimization and hyperband algorithm, leave-one-out method, back propagation neural network

摘要： 表面粗糙度和拉伸强度是衡量熔融沉积制造(FDM)成型件质量的重要指标,但由于FDM工艺参数众多,且与FDM 成型件质量之间呈现非线性关系,因此传统方法难以准确预测这两项指标。为此,提出一种贝叶斯超频道优化算法(BOHB)与BP神经网络相结合的FDM 3D打印成型件质量预测方法以提高预测精度与稳定性。将层厚、扫描次数和填充间隔这三个工艺参数作为模型的输入;利用BOHB算法对BP神经网络的超参数进行优化得到BOHB-BP模型;使用中心复合实验获取表面粗糙度和拉伸强度的实验数据,在以上两种数据集上根据留一法验证模型的精度与稳定性;将模型BOHB-BP与模型GA-BP和BP的预测情况进行对比实验,证明了所提方法在不同数据集上均有更好的预测精度与稳定性。

关键词: 熔融沉积制造, 质量预测, 贝叶斯超频道优化算法, 留一法, BP神经网络