基于深度学习的完全填充型熔融沉积成型零件质量预测方法

Quality prediction method of fully filled fused deposition molding parts based on deep learning

抗拉强度、翘曲度和表面粗糙度是衡量熔融沉积成型(FDM)零件质量的重要指标,对其准确、稳定的预测有助于FDM工艺的发展。为此,提出一种基于优化深度信念网络的完全填充型FDM零件质量预测方法。首先根据FDM的生产工艺选取影响零件质量指标的主要变量,利用相关性分析方法确定对产品质量影响最大的工艺参数组合,以获取预测模型的输入变量;其次以10—折交叉验证的验证误差作为适应度值,基于网格搜索确定稀疏约束深度信念网络(SDBN)的最佳超参数组合,采用自适应布谷鸟搜索(ACS)算法对SDBN进行优化,构建完全填充型FDM零件质量预测模型;最后,将所提的ACS-SDBN与ACS-DBN、深度信念网络(DBN)和BP的预测结果进行对比,结果表明基于ACS-SDBN模型的完全填充型FDM零件质量预测方法具有更好的预测精度和稳定性。

Tensile strength, warping degree and surface roughness are important indicators to evaluate the quality of Fused Deposition Modeling(FDM) parts. Accurate and stable prediction of the fused deposition modeling is conducive to the development of FDM technology. For this reason, a quality prediction method for fully filled FDM parts based on optimized deep belief network was proposed. According to the FDM production process, the main variables affecting the quality of parts were selected, and the process parameter combination with the greatest influence on the product quality was determined by using the correlation analysis method, so as to obtain the input variables of the prediction model. The validation error of 10-fold cross-validation was taken as the fitness value, and the optimal hyperparameter combination of Sparse Deep Belief Network(SDBN) was determined based on the grid search. Adaptive Cuckoo Search(ACS) algorithm was used to optimize SDBN, and a quality prediction model for fully filled FDM parts was established. The prediction performance of ACS-SDBN model was compared with that of the ACS-DBN, DBN and BP models, and the results showed that the fully filled FDM parts quality prediction method of sparse constraint Deep Belief Network optimized by Adaptive Cuckoo Search algorithm had better prediction accuracy and stability.