基于机器学习的RDX-CMDB推进剂安全性能预测  

Machine-learning prediction for safety of RDX-CMDB propellants

作　　者：郭延芝[1] 吴艳玲 刘润青 赵凤起[3] 徐司雨[3] 蒲雪梅[1] GUO Yan-Zhi;WU Yan-Ling;LIU Run-Qing;ZHAO Feng-Qi;XU Si-Yu;PU Xue-Mei(College of Chemistry,Sichuan University,Chengdu 610064,China;School of Environment and Safety Engineering,North University of China,Taiyuan 030051,China;National Key Laboratory of Energetic Materials,Xi'an Modern Chemistry Research Institute,Xi’an 710065,China)

机构地区：[1]四川大学化学学院,成都610064 [2]中北大学环境与安全工程学院,太原030051 [3]西安近代化学研究所含能材料全国重点实验室,西安710065

出　　处：《四川大学学报(自然科学版)》2025年第1期228-236,共9页Journal of Sichuan University(Natural Science Edition)

基　　金：国家重大专项(JG2021362);国家自然科学基金(62475177)。

摘　　要：为了满足对于固体推进剂安全性能预测的迫切需求,本研究提出了一种基于机器学习的推进剂安全性能预测方法.针对现有推进剂的安全性能数据量不足的问题,本研究引入主动学习混合插值(AL-Mixup)方法对数据进行了增强.综合考虑推进剂原料的组分种类、含量和颗粒粒度对安全性能的影响,本研究构建了含RDX改性双基(RDX-CMDB)推进剂的摩擦感度与撞击感度预测模型.以28组实验测定获得的RDX-CMDB样本为初始数据集,本研究获得了不同数量的增强样本.经Z-score特征标准化处理后,本研究用10种不同的机器学习算法构建了多种预测模型.以R^(2)、RMSE和MAE作为评价指标,本研究利用十折交叉验证法对比分析了不同数量增强样本及不同机器学习算法的组合的预测能力,最终得到两个性能优异的预测模型,一个是基于支持向量回归算法的摩擦感度预测模型(R^(2)=0.7950),另一个是基于人工神经网络算法的撞击感度预测模型(R^(2)=0.8932).对外部样本的测试结果显示,两个模型的预测精度均超过88%,同样令人满意.本研究首次实现了RDX-CMDB推进剂安全性能的快速预测,可望为此类推进剂的安全使用和配方优化提供理论依据.This paper focuses on the quantitative prediction of safety of solid propellants based on machinelearning(ML)algorithm.An active-learning based Mixup(AL-Mixup)algorithm is introduced to augment the insufficient valid safety data of solid propellants.The effects of raw material composition,content and particle size on the safety of solid propellants are comprehensively considered.28 experimental samples of RDXCMDB propellants obtained from actual experiments are taken as the initial dataset,which is characterized as feature vectors based on the component content information and particle size value.This dataset is then augmented by using the AL-Mixup algorithm,different level samples are obtained.After Z-score feature standardization,10 ML algorithms,including the multiple linear regression(MLR),partial least-square regression(PLSR),kernel ridge regression(KRR),least absolute shrinkage and selection operator(LASSO),K nearest neighbors(KNN),support vector regression(SVR)based on different kernels,random forest(RF),extreme gradient boosting(XGB),light gradient boosting machine(LGB)and artificial neural network(ANN),are used to construct prediction models.Comparisons is done for the combinations of augmented data and ML algorithms by taking the coefficients of determination(R^(2)),root mean square error(RMSE)and mean absolute error(MAE)as evaluation metrics,and the 10-folds cross-validation method is used as the evaluation methodology.Two best safety prediction models,including the SVR(rbf)model for friction sensitivity of RDX-CMDB propellants with R^(2) of 0.7950 and the ANN model for impact sensitivity of RDX-CMDB propellants with R^(2) of 0.8932,are obtained.Finally,in order to verify the practical prediction ability of the two models,4 external samples that are not involved in the training process of models are predicted.Relative errors between the experimental and predicted values are calculated,respectively.It is shown that the two models also exhibit excellent prediction ability with accuracies higher than 88%,

关 键 词：RDX-CMDB推进剂 摩擦感度 撞击感度 机器学习 数据增强 

分 类 号：O29[理学—应用数学]