不平衡工艺参数数据集的高温透平叶片铸件质量预测方法

Quality Prediction Method of High Temperature Turbine Blade Castings Based on Unbalanced Process Parameter Data Set

针对熔模精密铸造工艺参数数据集射线检测(RT)结果存在合格与不合格数量严重不平衡问题,提出一种基于概率分布的合成少数类集成学习(SyMProD-Stacking)的铸件质量预测方法。该方法首先对原始数据集进行预处理以保证数据质量,然后利用Z分数去除噪声数据,为每个少数类实例(不合格铸件)分配一个概率并基于此概率分布生成样本数据以获取平衡数据集,利用极端梯度提升模型(XGBoost)对所有工艺参数特征进行重要性排序并剔除部分排名靠后的工艺参数,最后将轻量级梯度提升机(LightGBM)、随机森林(RF)、支持向量机(SVM)和XGBoost模型进行Stacking集成并利用平衡数据集构建质量预测模型。以高温透平叶片制造过程精铸工艺为例,对所提出的质量预测方法进行验证,结果表明:相比于原始数据集构建的预测模型,利用了SyMProD过采样方法构建的预测模型不合格铸件的预测准确率提升了75.4%;相比于单一算法模型,所提质量预测方法的曲线下面积(A_(AUCROC))、几何均值(G_(m))以及F_(1)分数(F_(1))这3项性能指标分别提升了5.48%~11.59%、3.78%~8.92%、5.72%~11.39%,所提出的方法能够很好地预测高温透平叶片精铸过程在不平衡问题下的铸件质量。

In response to the significant imbalance between the quantities of qualified and non-qualified results in the radiographic testing(RT)of the investment precision casting process parameters dataset,this paper proposes a casting quality prediction method using SyMProD-Stacking ensemble learning.The method begins by preprocessing the original dataset to ensure data quality.It then employs Z-scores to eliminate noisy data,assigns a probability to each minority class instance(non-qualified castings),and generates sample data based on this probability distribution to obtain a balanced dataset.XGBoost is used to rank the importance of all process parameter features and removes some of the lower-ranking parameters.Finally,the LightGBM,RF,SVM,and XGBoost models are stacked together through ensemble learning,and a quality prediction model is constructed using the balanced dataset.Taking the precision casting process in the manufacturing of high-temperature turbine blades as an example,the proposed quality prediction method is validated.The results indicate that the predictive model constructed using the SyMProD oversampling method significantly outperforms the model built from the original dataset,with a 75.4%improvement in the accuracy of predicting non-qualified castings.Stacking ensemble learning,compared to individual algorithm models,achieves improvements of 5.48%—11.59%,3.78%—8.92%,and 5.72%—11.39%in terms of area under curve,geometric mean,and F_(1),respectively.