主管：中华人民共和国应急管理部
主办：应急管理部天津消防研究所
ISSN 1009-0029  CN 12-1311/TU
消防科学与技术

消防科学与技术 ›› 2023, Vol. 42 ›› Issue (5): 609-614.

• • 上一篇    下一篇

基于BP人工神经网络的有机化合物爆炸下限预测

时静洁1, 赵薇1, 陈小林2, 陈常豪2   

  1. （1. 常州大学 安全工程学院，江苏 常州 213164； 2. 天地（常州）自动化股份有限公司，江苏 常州 213015）
  • 出版日期:2023-05-15 发布日期:2023-05-15
  • 作者简介:时静洁（1988- ），女，江苏常州人，常州大学安全与工程学院讲师，博士，主要从事化工工艺热安全分析方面的研究，江苏省常州市武进区滆湖中路1号，213164。
  • 基金资助:
    江苏省高等学校自然科学研究项目（19KJB620002）；常州市科技支撑计划（社会发展）项目（CE20205019）

Predicting lower explosive limit of organic compounds based on BP artificial neural network

Shi Jingjie1, Zhao Wei1, Chen Xiaolin2, Chen Changhao2   

  1. (1. School of Safety Engineering, Changzhou University, Jiangsu Changzhou 213164, China;2. Tiandi (Changzhou) Automation Co., Ltd., Jiangsu Changzhou 213015, China)
  • Online:2023-05-15 Published:2023-05-15

摘要： 摘 要：运用定量结构-性质关系对458种有机化合物的爆炸下限展开预测研究。首先运用Dragon 2.1软件计算并预筛出708种分子描述符，随后采用遗传算法确定了5个特征分子描述符作为模型的输入变量，最后运用SPSS和MATLAB分别构建了多元线性回归线性模型、支持向量机与人工神经网络两种非线性模型。研究结果为：MLR模型的训练集和测试集的复相关系数R2分别为0.838 7和0.858 8；SVM模型的R2分别为0.856 9和0.877 9；ANN模型的R2分别为0.928 4和0.932 8。由此表明，无论是训练集还是测试集，SVM模型的预测效果均优于MLR模型，ANN模型的预测效果均优于SVM模型，有机化合物的爆炸下限与其分子结构之间存在着较强的非线性关系。此外，本研究采用内外验证方法及与其他研究的比较对模型性能进行了验证，证实了ANN模型对爆炸下限具有较好的预测能力。通过绘制Wiliams图分析了模型的应用域，验证了所建模型均具有良好的泛化能力和鲁棒性。通过QSPR方法预测有机化合物的爆炸下限，能为危险化学品的风险管控及安全工艺的研究提供有力的理论和技术支持。

关键词: 定量结构-性质关系, 人工神经网络, 多元线性回归, 支持向量机, 爆炸下限, 有机化合物

Abstract: This paper uses Quantitative Structure-Property Relationship correlation method to predict the Lower Explosive Limit experimental values of 458 organic compounds. First of all, Dragon 2.1 software was used to calculate and pre-screened 708 molecular descriptors. Then Genetic Algorithm was used to determine 5 characteristic molecular descriptors as the input variables of the model. Finally, SPSS software was used to construct the Multiple Linear Regression linear model, and MATLAB software was used to construct Support Vector Machine and Artificial Neural Network nonlinear models. The results showed that for MLR model, the complex correlation coefficient R2 of training and test set were 0.838 7 and 0.858 8 respectively. For SVM model, R2 is 0.856 9 and 0.877 9. For ANN model, R2 is 0.928 4 and 0.932 8. It showed that the prediction effect of SVM model was better than MLR model, and the prediction effect of ANN model was better than SVM model. There was a strong nonlinear relationship between the LEL of organic compounds and their molecular structure. In addition, the performance of the ANN model was evaluated by internal and external verification methods and compared with previous studies, which proved that the ANN model had the best prediction accuracy for the LEL. The Wiliams diagram was drawn to analyze the application domain of the model, and the model had well generalization ability and robustness. The LEL of organic compounds was predicted by the method of QSPR, which provided effective theoretical research and technical support for the risk control and safety process research of hazardous chemicals.

Key words: Quantitative Structure-Property Relationship, Artificial Neural Network, Multiple Linear Regression, Support Vector Machine, Lower Explosive Limit, organic compound