三篇硕士学位论文摘要

含能物质能量与安全性统计规律及综合效应研究与应用

硕士研究生：呙娓佽   导师：吴超

答辩时间：2012年11月24日

摘 要

近年来，随着我国国防科技工业、航天事业及民用领域的发展，对含能物质的需求激增。作为生产制造混合炸药、发射药、固体推进剂的基本材料，含能物质本身蕴含较大能量且分解具有爆轰效应，在生产、使用及储存过程中具有极大的危险性。因此随着含能物质应用范围的扩大，其安全性问题日益凸显。本研究在此背景下，探讨含能物质能量与安全性方面的之间的规律及综合效应，旨在保证含能物质安全性的前提下提高其能量利用率。

本论文在大量含能物质诸理化参数、性能参数等参量的文献调查和数据统计的基础上，以含能物质的能量与安全性为着眼点，运用理论分析、统计学、综合评价分析的原理和方法，首先进行含能物质性能和感度的单因素分析，对表征能量性的参量—爆速进行了参数关联和预估建模，同时对含能物质安全性表征—感度进行全面衡量。在此基础上探索了含能物质能量与安全性的转化关系，提出基于性能和感度的炸药综合评定方法。具体研究内容和结论如下：

（1）对含能物质能量危险性进行了阐述，综述了国内外含能物质研究发展状况及研究前沿，对含能物质能量与安全性方面的相关研究进行了文献探索。

（2）简介了含能物质的特点、分类，从物理性质、化学性质、热力学参数、动力学参数、爆轰性能及感度特性等方面罗列了含能物质诸多参数的概念及计算、测试方法，并通过文献调查和数据统计，得到国防工业和民用行业常用的65种含能物质的各类参数。

（3）从能量角度出发，应用参数关联法挖掘含能物质性能参数的相关性。散点图及相关系数研究表明含能物质性能参数—爆速D与氧平衡指数OB100及密度ρ的关联性较高；采用回归分析建立含能物质动力学参数前指因子lgA与活化能Ea的一元线性回归模型，证实了同类型含能物质在热分解过程中存在动力学补偿效应。在参数关联的基础上，应用多元线性回归法MLR及BP神经网络法BPNN建立炸药爆速的定量结构-性质相关研究QSPR预估模型，研究结果表明利用二种方法筛选出的描述符都能够反映爆炸特性，而应用神经网络算法能得到比传统的线性模型精度更高的预测结果。

（4）从安全性的角度出发，进行了含能物质安全性-综合感度特征值的探讨。研究基于BZA-1法，将不同试验条件下的感度数据分配权重，并通过模糊加权层次法得到常用太安等12种含能物质综合感度特征值的危险性排序。

（5）基于ISL试验数据值，将含能物质性能和感度两方面各3组试验值进行综合统计分析，每组选取最高试验数据值赋值为特征值1，其余物质试验值与此基准物质的比值作为其特征值，将两方面的数值综合为性能和感度方面的特征，并通过合成运算及作图法得到物质能量与安全性的转化关系。

（6）提出了含能物质性能及安全性的模糊动态聚类综合评价法。对通过文献收集到代表炸药能量和安全性能的各项评价指标进行动态聚类，寻找最佳阈值 ，灵活选择动态结果进行综合评判，以分析性能及安全性的优劣差异。

关键词：含能物质，能量，安全性，统计分析，参数关联，性能预估，综合评价

ABSTRACT

In recent years, the demands for energetic materials surge due to the development of the national defense industry, space undertakings and civil profession. As the basic ingredients of mixed explosives, pyrotechnics and propellants, energetic materials are possessed with high energy and detonation potential, thus they have great risks in the production, procession and storage process in accordance. With the expansion of their application, the safety issues of energetic materials become increasingly prominent. In this context, mutual discipline and comprehensive effect research on the energy and safety aspects of energetic materials are studied in this paper, aiming to improve energy utilization on the premise of ensuring safety.

On the basis of literature retrieval and data recollection of physical and chemical parameters and performance qualities, focusing on energy and safety aspects of energetic materials and by adoption of theoretical analysis, statistical principles and comprehensive evaluation methods, single factor analysis of performance and sensitivity of energetic materials is proceeded firstly, in which parameter correlation and prediction modeling studies are carried out and safety overall judgment is conducted as well. The conversion relationship is explored on that portfolio and the synthetic evaluation approach is proposed based on performance and sensitivity characteristics. Main concrete research contents and conclusions as listed as follows:

(1) Energy hazards of energetic materials are discussed in the beginning. Reviews on research status and advance at home and abroad are summarized and literature retrieval is conducted on energy and safety aspects of energetic materials.

(2) The characteristics and classification of energetic materials are briefly described. Physical attributes, chemical attributes, thermodynamic parameters, kinetic parameters, detonation qualities and sensitivity properties are introduced. By literature retrieval and recollection, the parameters of 65 energetic materials of different kinds are obtained.

(3) Correlation of performance qualities of energetic materials is explored by the application of parameter correlation method from the viewpoint of energy. Correlation scatter diagram and simple correlation coefficient demonstrate that detonation velocity D corresponds highly with oxygen balance index OB100 and density ρ. Regression analysis is applied to establish the linear regression model of the anaphoric factor lgA and activation energy Ea to prove the kinetic compensation effect of energetic materials of same type in their thermal decomposition process. Then the multiple linear regression and BP neutral network approach are utilized to establish the QSPR prediction model of detonation velocity repsctively. Results show that the selected descriptors of both models can reflect the explosion characteristics and BP neutral network exceeds the traditional linear model in prediction accuracy.

(4) Comprehensive sensitivity eigenvalue is analyzed from the viewpoint of safety. Based on the BZA-1 method, different weights are assigned to different sensitivity data on dissimilar test condition. Risk sort of comprehensive sensitivity eigenvalue of 12 energetic materials is acquired by fussy weighted hierarchical method.

(5) Experimental data from several performance and sensitivity tests have been combined after normalization to define a single performance and a safety characteristic term, respectively. This allows evaluating energetic compounds with regard to a well balanced ratio of performance and sensitivity. A graph shows an imaginary border line what has to be interpreted in the sense that high performance is accompanied by an enhanced sensitivity and that an insensitive explosive will not exhibit a top performance.

(6) Fussy dynamic clustering evaluation method is proposed to comprehensively assess performance and sensitivity characteristics of energetic materials. Multiple energy and safety indexes are selected for dynamic clustering and optimal threshold value are chosen to flexibly select dynamic results to roundly evaluate the discrepancy of performance and safety of energetic materials.

KEYWORDS: energetic materials, energy, safety, statistical analysis, parameter correlation, performance estimation, comprehensive evaluation

QSPR/QSAR在有机物危险特性预测中的应用

硕士研究生：李冀    导师：吴超

答辩时间：2012年11月24日

摘 要

近年来，随着我国经济和化工产业的飞速发展，越来越多的危险化学品出现在生产、经营、运输及使用中，这无疑对人类和社会带来了潜在的威胁，因此对化学物质进行危险性评价显得越来越重要。理化性质是评价化学品危险性的重要指标，但由于各种原因，目前还没有一个完整的数据库。定量结构-性质/活性关系(Quantitative Structure-Property/Activity Relationship, QSPR/QSAR)方法的出现为化学品的危险性预测提供一个可靠的手段。一旦建立了可靠的模型，既可以用它来预测新的甚至是尚未合成的化合物的各种性质，而且可以在微观上了解分子结构对性质的影响，这对新分子的设计有一定的指导作用。

本论文利用遗传函数算法(Genetic Function Approximation, GFA)来选择描述符，用多元线性回归(Multivariable Linear Regression, MLR)的方法建立线性模型；随后还使用BP神经网络(Back-Propagation Network, BPNN)和支持向量机(Support Vector Machines, SVM)来建立非线性模型，得到的结果令人满意。具体可分为以下几个方面的内容：

（1）阐述了QSPR/QSAR的基本原理、研究步骤以及研究进展；详细的解释了BPNN和SVM的基本原理。

（2）建立了一个用于研究脂肪族化合物的急性毒性的QSAR模型。利用GFA选择分子描述符，分别用MLR和BPNN建立急性毒性与分子描述符之间的线性和非线性模型。测试集中复相关系数R2分别为0.760和0.814，平均绝对误差AAE为0.314mmol/L和0.296mmol/L，表明非线性模型的拟合度和预测精度均优于线性模型。该方法提供了一个基于分子结构预测脂肪族化合物急性毒性的新途径。

（3）预测了结构类型互不相同的1056种有机物的燃烧下限。通过GFA筛选出4个与燃烧下限密切相关的结构参数。MLR和BPNN分别用来建立线性和非线性模型。测试集中复相关系数R2分别为0.956和0.978，均方根误差RMSE分别为0.107vol%和0.077vol%。结果表明，BPNN模型性能优于MLR。

（4）运用MLR和SVM方法，分别建立了91个脂肪醇化合物的结构与其闪点之间的线性和非线性的QSPR模型。测试集中复相关系数R2分别为0.976和0.979，平均绝对误差AAE为2.870K和2.706K。结果表明，通过GFA筛选出的3个描述符能很好的表征脂肪醇化合物的闪点。

（5）将QSPR方法应用于液态烃化合物定量结构性质关系的研究。应用MLR和SVM建立了液态烃燃烧热与三个描述符的相关性模型。测试集中的复相关系数R2分别为0.992和0.993，平均绝对误差AAE为121kJ/mol和88kJ/mol。该方法的提出为工程上预测液态烃化合物燃烧热提供了有效途径。

关键词 化学信息学；定量结构性质/活性关系；遗传函数算法；BP神经网络；支持向量机

ABSTRACT

In recent years, with the rapid development of China's economy and the chemical industry, a growing number of hazard chemicals appear in the production, management, transport and usage process, which will undoubtedly bring potential threats to human and society. Therefore it becomes increasingly important to assess the hazards of chemicals. Physical-chemical properties are essential attributes for the evaluation of hazard chemicals. Due to various reasons, there is no such complete database to record the chemicals’ physical-chemical properties. Quantitative Structure-Property/Activity Relationship (QSPR/QSAR) provides a reliable means for assessing the hazard chemicals. Once a reliable model is established, it can then be used to predict various properties of new materials and even compounds yet not synthesized, and micro molecular structure nature can be understood as well. It is constructive to the design of new molecules.

In this study, molecule descriptors are selected by the genetic function algorithm (GFA), using multiple linear regression (MLR) method to create a linear model; then non-linear model are built by BP neural network (BPNN) and support vector machines (SVM), and the results are satisfactory. Concrete main contents are listed as follows:

The first chapter presents the basic principles of QSPR/QSAR, research steps and research progress, and detailed explanation of the basic principles of BPNN and SVM are introduced.

The second chapter establishes a QSAR model which is used to study the acute toxicity of fatty compounds. Molecular descriptors are selected by GFA, using the MLR and BPNN to establish the linear and nonlinear models of acute toxicity and molecular descriptors. Multiple correlation coefficient (R2) of the test set are 0.760 and 0.814, and the average absolute error (AAE) are 0.314mmol/L and 0.296mmol/L respectively. Results show that the non-linear model fitting and forecasting accuracy are better than the linear model. The method provides a way to predict the acute toxicity of fatty compounds based on the molecular structure.

The third chapter predicts lower flammable limit of 1056 different kinds of organic compounds. Four structural parameters closely related to the lower flammable limits are seletced through GFA. MLR and BPNN method are used to establish the linear and nonlinear models. Multiple correlation coefficient (R2) of the test set are 0.956 and 0.978, respectively, the root mean square error (RMSE) are 0.107vol% and 0.077vol% respectively. The results show that the BPNN model performance is better than the MLR.

The fourth chapter establishes the relationship of flash point and molecular structure of 91 fatty alcohol compounds by MLR and SVM methods. In the test set, multiple correlation coefficients (R2) are 0.976 and 0.979, and the average absolute error (AAE) are 2.870K and 2.706K. The results show that three descriptor selected through GFA can be a good characterization of the flash point of the fatty alcohol compounds.

In the fifth chapter, QSPR method is used to study the quantitative structure-property relationship of liquid hydrocarbons. Correlation model of heat combustion with three descriptors of liquid hydrocarbons is established by application of MLR and SVM approach. The multiple correlation coefficient (R2) of the test sets are 0.992 and 0.993, and the average absolute error (AAE) are 121kJ/mol and 88kJ/mol respectively. This method provides an effective way to predict the heat combustion of liquid hydrocarbon.

Key words: chemoinformatics; quantitative structure-property/activity relationship; genetic function approximation; back-propagation network; support vector machines

有机过氧化物的参数统计分析及安全性研究

硕士研究生：谭波    导师：吴超

答辩时间：2012年11月24日

我国安全学科近几年发展较快，2009年国家标准学科分类体系GB/T13745-2009《学科分类与代码》一级学科“安全科学技术”新增5个二级学科，其中就有代码为62023的“安全物质学”，其研究还处在初步探索阶段。本文通过选取危险化学品中的有机过氧化物类物质，对其安全因子的相关性进行研究，并探索“安全物质学”的研究方法和实证例子。

本文通过搜集和查阅相关文献，在对国内外安全物质学和有机过氧化物研究现状进行系统综述的基础上，统计有机过氧化物的各项参数数据，采用相关性分析、因子分析、聚类分析、判别分析和Copeland计分排序法，对有机过氧化物的参数数据进行分析，得到了影响有机过氧化物安全性的5个公因子和基于其安全性的综合分类结果。针对影响其安全性最主要的参数，采用基团贡献法、BP神经网络、多元线性回归、偏最小二乘法和支持向量机进行物性参数预测。本论文主要研究内容和成果如下：

（1）将所收集统计的72种有机过氧化物的参数数据分成三类：分子成分参数、分子结构参数和物性参数。分别对三类数据进行相关性分析，并结合参数定义删除部分参数。

（2）对相关性分析筛选出的72种有机过氧化物的23个参数指标进行因子分析，提取出5个公因子并分别定义为：燃烧爆炸因子、分解因子、扩散因子、污染性与腐蚀性因子、毒害性因子。与以往研究相比较，结果显示5个公因子可以很好的表征有机过氧化物的安全性。

（3）根据因子分析的结果及物质统计数据的完整性，选取33种有机过氧化物及其10个参数。先对其中28种有机过氧化物进行聚类分析，划分为3类。对5种未知类别的有机过氧化物进行判别分析。再对33种有机过氧化物进行Copeland计分排序。最后将三种方法的分类结果进行比较，判别分析与聚类分析的分类匹配率达到100%，Copeland计分排序的结果与前两者的匹配率也达到了54%。

（4）依据燃烧爆炸因子最能代表物性参数闪点，设计了一种基于基团贡献法BP神经网络预测有机过氧化物闪点的方法。将66种有机过氧化物中的50种作为训练集16种为预测集进行闪点预测。结果表明，闪点的预测值与文献值符合良好。该预测方法不需计算参数，且结果可靠。

（5）根据前文结合文献研究可知有机过氧化物的反应危害主要取决于起始分解温度和分解热。选取14种有机过氧化物作为训练样本，3种为预测样本，确定13个参数为描述符，分别采用多元回归分析、偏最小二乘法和支持向量机进行起始分解温度和分解热的预测，并比较和评价三种方法所得模型的有效性。最终确定PLS和SVM分析得到的预测模型与文献数据有较高的线性拟合且具有良好的预测能力。

关键词：有机过氧化物，安全性，因子分析，分类研究，BP神经网络，偏最小二乘法，支持向量机

ABSTRACT

Safety science has been developing rapidly for many years in China. The GB/T13745-2009 “subject classification and code” expanded five grade-2 subjects which include safety material science. The research of safety material science is still in the initial exploratory stage. This paper selected a series of organic peroxides as object to explore the research contents and methods of safety material science.

Based on the collection and access to relevant literature, systematic review was given to the research status of the safety material science and organic peroxides at home and abroad. The parameter data of organic peroxides were count and analyzed by correlation analysis, factor analysis, cluster analysis, discriminant analysis and Copeland score sort. Five common factors affecting the safety of organic peroxides and classification results based on their safety were obtained. Then the main physical parameters were predicted by group contribution method, BP neural network, multiple linear regression, partial least squares and support vector machine. The main research and conclusions in this paper are as follows:

(1) The statistics parameter data of 72 organic peroxides were divided into three categories: molecular composition parameter, molecular structure parameter and physical parameter. Through the correlation analysis among three categories data and the definition of the parameter, few parameters were deleted.

(2) By the factor analysis of 23 parameters of 72 organic peroxides, five common factors were extracted and respectively defined as combustion and explosion, decomposition, diffusion, pollution and corrosion, and toxicity. Compared to previous studies, the result showed that the five common factors can be a good characterization to the safety of organic peroxides.

(3) Based on the result of factor analysis and the integrity of the statistical data, 33 organic peroxides and their 10 parameters were selected. 28 organic peroxides were classified as three categories by the cluster analysis. 5 unknown types’ organic peroxides were analyzed by discriminant analysis. Then 33 organic peroxides were sort by Copeland score. The comparison among the results of the three methods demonstrated that there was high consistent.

(4) According to combustion and explosion can best represent flash point, a method based on BP neural network and group contribution method to predict the flash point of organic peroxides was designed. 50 organic peroxides were chosen as training set and 16 organic peroxides were chosen as testing set. The result showed that the predicted values were in good agreement with literature values. The prediction method didn’t require calculation of the parameters, and the results were reliable.

(5) According to previous research, it was known that the reaction hazards of organic peroxides depend on the initial decomposition temperature and heat of decomposition. 14 organic peroxides were chosen as training set, 3 organic peroxides were chosen as testing set, and 13 parameters were determined as the descriptors. The initial decomposition temperature and heat of decomposition were predicted by multiple linear regression, partial least squares and support vector machine. The effectiveness of three methods were compared and verified. Finally, the predicted models of PLS or SVM had a higher linear fit with literature data and a good predictability.

KEY WORDS: organic peroxides, safety, factor analysis, classification, BP neural network, partial least squares, support vector machine