预测改性双基推进剂燃速的机器学习建模  

Machine learning modeling for predicting burning rate of CMDB propellants

作　　者：陈少臣 郭建雄 熊思璇 刘哲人 王晓晨 马煜[1] 代志龙[1] CHEN Shaochen;GUO Jianxiong;XIONG Sixuan;LIU Zheren;WANG Xiaochen;MA Yu;DAI Zhilong(Hubei Institute of Aerospace Chemotechnology,Xiangyang 441003,China)

机构地区：[1]湖北航天化学技术研究所,襄阳441003

出　　处：《固体火箭技术》2025年第1期124-132,共9页Journal of Solid Rocket Technology

摘　　要：为高效评估改性双基(CMDB)推进剂配方的燃速,使用机器学习(ML)方法建立CMDB推进剂配方的燃速预测模型。首先,对含有137个样本的CMDB推进剂配方-燃速数据集进行相关性分析以确定ML模型的17个输入特征,随后使用该数据集训练、优化与评估6个ML模型,包括岭回归(RR)、随机森林、梯度提升树(GBDT)、极端梯度提升(XGB)、支持向量机(SVM)和人工神经网络(ANN);然后,采用置换特征重要度与沙普利加性(SHAP)解释方法计算预测精确度最高的黑盒模型的特征重要度,并使用SHAP寻找特征与燃速之间的关系;最后,计算白盒模型RR的权重系数,以获取其特征重要度、特征与燃速的关系,并与前一个模型的结果进行对比。结果表明,GBDT、XGB、SVM和ANN模型的决定系数(R^(2))均超过了0.99,其预测值的残差和相对百分误差主要分布在-1~1 mm·s^(-1)、-10%~10%。通过解释SVM(R^(2)排名第一)与RR模型,发现两个模型都为燃速抑制剂、Al、催化剂的质量分数和工作压强赋予了很高的特征重要度,会显著影响燃速。此外,增加燃速抑制剂等组分的质量分数会降低燃速,而增加Al、催化剂等组分的质量分数与工作压强则会提升燃速。To efficiently evaluate the burning rate of CMDB propellant formulations,the machine learning(ML)method was used to establish the burning rate prediction models of CMDB propellant formulation.Firstly,a correlation analysis was conducted on a dataset containing 137 samples with CMDB propellant formulation-burning rates to confirm the 17 input features of the ML models;subsequently,six ML models were trained,optimized,and evaluated using this dataset,including ridge regression(RR),random forests,gradient boosting decision tree(GBDT),extreme gradient boosting(XGB),support vector machine(SVM)and artificial neural network(ANN).Subsequently,the feature importance of the black box model with the highest prediction accuracy was calculated by employing interpretation methods of permutation feature importance and shapley additive explanations(SHAP),and the relationship between the feature and burning rate was found by using SHAP.Finally,the weight coefficient of the white box model RR was calculated to obtain its feature importance,the relationship between features and burning rate,and the results were compared with the previous model.The results show that the coefficient of determination(R^(2))of GBDT,XGB,SVM,and ANN are all higher than 0.99,with their prediction residuals and relative percentage errors primarily distributed within-1~1 mm·s^(-1)and-10%~10%,respectively.By explaining the SVM model(ranked first in R^(2))and the RR model,it is found that both models assign high feature importance to the mass fraction of burning rate inhibitors,Al,catalysts,and working pressure,which significantly affect the magnitude of burning rate.In addition,increasing the mass fraction of components such as burning rate inhibitors will reduce the burning rate,while increasing the mass fraction of components such as Al and catalysts,and working pressure will increase the burning rate.

关 键 词：CMDB推进剂 燃速 机器学习 置换特征重要度 沙普利加性解释 

分 类 号：V512[航空宇航科学与技术—航空宇航推进理论与工程]