当前位置：首页 > 过刊浏览->2025年46卷第12期
基于数据增强与机器学习的电子束熔化增材制造 TiAl-4822 合金疲劳寿命高精度预测研究
High-accuracyFatigue Life Prediction of Electron Beam Melting Additively ManufacturedTiAl-4822 Alloy Based on Data Augmentation and Machine Learning
浏览(17) 下载(1)
- DOI:
- 作者:
- 叶嘉峰 1，林博超 2，鲍伊达 3，陈玮 2
YE Jiafeng1,LIN Bochao2,BAO Yida3,CHEN Wei2
- 作者单位:
- 1. 上海交通大学 材料科学与工程学院，上海 200240；2. 中国航空制造技术研究院，北京 100095；3. 威斯康星大学斯托 特分校，科学、技术、工程、数学与管理学院，威斯康辛州 梅诺莫尼市 54751
1. School of Materials Science and Engineering, Shanghai Jiao Tong University, Shanghai 200240, China; 2. AVIC Manufacturing Technology Institute, Beijing 100095, China; 3. University of Wisconsin-Stout, College of Science, Technology, Engineering, Mathematics & Management, Menomonie 54751, USA
- 关键词:
- 钛铝合金；电子束熔化；疲劳寿命；数据增强；机器学习；分层神经网络；稳健回归
TiAl alloy; electron beam melting; fatigue life; data augmentation; machine learning; hierarchical neural network; Huber regression
- 摘要:
- 电子束熔化(electron beam melting, EBM)增材制造制备的 Ti-48Al-2Cr-2Nb(TiAl-4822)在复杂服役条件下疲劳寿命波动显著，影响工程可靠性。 基于包含 103 个 EBM 制备 TiAl-4822 件的疲劳测试数据，本文结合数据增强与机器学习(层级神经网络、稳健回归)构建高精度疲劳预测模型(整体误差小于 20%)。 该模型首先使用合成少数类过采样技术(SMOTE)增强并且平衡数据，结合分层神经网络分类模型的判别是否通过疲劳测试(分类准确率达 80%)；随后，针对未通过疲劳测试件，采用基于高斯噪声的回归型合成过采样方法(SMOGN)进行数据增强，结合稳健回归与分层神经网络组成的二阶模型进行疲劳寿命的回归预测(R2 达到 0.81，平均百分比误差为 7.3%)。 基于 Shapley 值的模型可解释性方法(SHAP)分析，加载频率、最大应力、温度及应力幅为主要影响因素。 研究建立了适用于小样本条件下的 EBM 制备的 TiAl-4822 服役疲劳寿命预测方法，为该合金在工程服役中的可靠性评估与优化设计提供了可行路径。Ti-48Al-2Cr-2Nb (TiAl-4822) materials fabricated via electron beam melting (EBM) additive manufacturing exhibit pronounced variations in fatigue life under complex service conditions, which affects their engineering reliability. To this end, on the basis of fatigue test data consisting of 103 EBM-fabricated TiAl-4822 samples, a high-accuracy fatigue life prediction model (overall error <20%) was developed by combining data augmentation techniques (SMOTE, SMOGN) with machine learning methods (hierarchical neural network, abbreviated as HNN, and Huber regression). The model first employs SMOTE to balance and argument the dataset and then integrates an HNN classifier to determine whether a sample would pass the fatigue test, achieving a classification accuracy of 80%. For the samples that failed the fatigue test, SMOGN was applied for data augmentation, and a two-stage model combining Huber regression with the HNN was used for fatigue life prediction, leading to an R2 of 0.81 and a mean absolute percentage error of 7.3%. SHAP analysis based on this model indicates that frequency, maximum stress, temperature, and stress amplitude are the primary influencing factors. A fatigue life prediction approach suitable for small-sample scenarios of EBM TiAl-4822 under service conditions is finally established.