传统的力学模型与新兴的数据驱动模型目前广泛用于增材制造的疲劳寿命预测中.其中,以连续损伤力学(Continuum Damage Mechanics,CDM)为代表的传统模型寿命预测模型存在着精度低、适用范围有限等问题,而以神经网络(Artificial Neural Network,ANN)为代表的数据驱动模型则始终受限于小样本工况.为解决上述问题,融合物理知识和数据信息的知识数据双驱动模型近年来发展迅速.以此类双驱动模型为研究目标,本文以激光粉末床熔融成型(Laser Powder Bed Fusion,LPBF)AlSi10Mg合金为研究对象,构建了可自动标定的CDM模型,并将其与基于ANN的数据驱动模型在各种工况下进行了结合,再进一步通过特征融合、参数融合和输出融合方法的手段,构建了三类以CDM模型与ANN模型为基础的知识-数据双驱动模型,并量化分析了它们在预测精度和数据需求等方面的性能.研究结果表明:基于参数融合的模型,训练数据修正作用较为显著,在预测精度方面受CDM模型影响最小,并在CDM模型拟合结果较差时也能确保一定精度;基于特征融合的双驱动模型能最大化利用CDM模型中的物理信息,在数据充足时具有最高的预测精度与稳定性;基于输出融合的模型以CDM模型的结果为主导,利用ANN进行修正,具有五种模型中最好的非训练域(外推)预测性能.这些结果对于进一步发展知识-数据双驱动的高精度增材制造疲劳寿命预测模型具有重要的参考价值.
Data-driven and Continuum Damage Mechanics-based Approach for Predicting Fatigue Lifein Additive Manufacturing
Additive manufacturing(AM)techniques have attracted widespread attention in aerospace and biomedical fields due to advantages like high material utilization and extensive design flexibility.How-ever,process-induced defects in AM-built components pose significant challenges for evaluating fatigue performance.The AM-built components are subjected to complex alternating loads in service,making it imperative to develop accurate fatigue life prediction models.Currently,two main approaches are widely employed:theoretical analysis and data-driven methods.Traditional life prediction models like continuum damage mechanics(CDM)suffer from limitations such as low accuracy and restricted applicability.Con-versely,data-driven models,such as artificial neural networks(ANN),encounter constraints when deal-ing with limited sample sizes.To address these issues,knowledge-data hybrid models have emerged as a promising approach that combines physical principles with data insights.In view of this,this study has de-veloped a calibrated CDM model and seamlessly integrated it with an ANN-based data-driven model.Em-ploying methods of feature,parameter,and output fusion,three types of hybrid models based on CDM and ANN have been developed.To quantitatively analyze the prediction accuracy and data requirements of these models,calculations using fatigue data obtained from laser powder bed fusion(LPBF)-processed Al-Si10Mg alloy have been performed.The results highlight the crucial role played by the corrective function of training data in the parameter fusion-based model,while indicating a relatively minor influence from the CDM model in terms of prediction accuracy.Moreover,this model retains a commendable level of accuracy even with suboptimal fitting outcomes from the CDM model.The hybrid model,which leverages feature fusion,maximizes the utilization of physical information from the CDM model,thus achieving the highest prediction accuracy and stability when ample data are available.The model based on output fusion,prima-rily guided by results of the CDM model and enhanced by ANN adjustments,demonstrates relatively supe-rior predictive capabilities in domains outside of the training set compared to other models.These findings provide significant reference value for the further development of high-accuracy,knowledge-data hybrid fa-tigue life prediction models in AM.
万方数据
维普
