【作者】 陈刚；

【导师】 杜树新；

【作者基本信息】 浙江大学 ， 模式识别与智能系统， 2010， 硕士

【摘要】 氟石粉在海运途中因运输水分含量过高会发生液化从而发生沉船事故。随着越来越多此类沉船事故的发生,人们意识到探讨氟石粉液化成因的重要性。在国内外对氟石粉液化研究较少的情况下,本文从考虑所有可能影响氟石粉液化的因素出发,利用特征选择方法筛选出影响氟石粉液化的主要因素并建立相应液化模型,为进一步机理模型建立及氟石粉厂家生产提供参考。主要工作如下:(1)为了寻找主要影响氟石粉液化的因素,先考虑了所有可能影响氟石粉液化的因素,按照相关实验标准进行实验并采集实验数据。总共采集了196个样品的20个属性的有效实验数据。(2)提出了基于回归预测误差的异常样本逐次剔除方法。该算法相比于一次性删除所有异常样品的剔除方法,给了被鉴定为异常样品二次确认的机会,这样会在很大程度上避免将一些正常样品误当作异常样品删除,同时又能达到提高回归预测模型精度的效果。(3)提出了基于回归预测误差和遗传算法集成的特征选择方法。该算法将所有的属性编码为一个遗传个体,利用回归预测模型的预测误差及个体的属性个数来评价该个体的适应度,通过选择、交叉及变异过程,不断繁殖与迭代,最终会收敛到一个最优的个体,此个体所包含的属性即为最优属性集合。同时在遗传算法中,提出了综合预测误差及属性个数的适应度函数确定方法,在选择算子与变异算子中引入了模拟退火算法的思想,使选择算子与变异算子得到改善,更有利于算法的寻优,加强了算法的全局搜索能力。(4)利用本文所提出的特征选择方法对实验数据进行分析得到包含8个属性的最优属性集,并对这8个属性在回归预测模型上进行灵敏度分析,指出这8个属性对氟石粉液化的影响。更多还原

【Abstract】 Fluoride powder containing large amounts of water is inclined to be liquefied, which will lead to the shipping wreck. As more and more accidents happen, researchers begin to investigate the causes of fluoride powder liquefaction. In the case of the few studies on the fluoride powder liquefaction both in research and industry domains, this thesis takes the initiate to study this topic. Firstly, the thesis considers all possible factors that affect the liquefaction of the fluoride powder. Then the feature selection method for selecting the main factors of the fluoride powder liquefaction is proposed, and an intelligent model of fluoride powder liquefaction is established. The main contributions of this thesis are provided as follows,(1) In order to find the main factors that affect the fluoride powder liquefaction, the thesis considers all possible factors that affect the liquefaction of fluorine powder. Some experiments under the relevant experimental standard are conducted with collecting 196 samples, each of which contains 20 properties for further investigation.(2) The progressive abnormal sample deletion method based on regression forecasting error is presented. Compared to the conventional abnormal sample deletion methods, the proposed algorithm gives the samples identified as abnormality a second testing opportunity, which could avoid the incorrect deletion of some normal samples as abnormal in large part. Meanwhile, this algorithm can dramatically improve the accuracy of regression forecasting model results.(3) The feature selection method based on regression forecasting error and genetic algorithm is presented. Processes of the algorithm are described as follows: Firstly, all of the properties of the fluoride powder liquefaction are encoded as a genetic entity. Secondly, this thesis evaluates the fitness of the individual using a fitness function based on regression forecasting error and the number of the properties belonging to this individual. Finally, an optimal individual will be selected through the repeated processes of selection, crossover and mutation. The properties of this individual are the prominent properties which affect the fluoride powder liquefaction. In the genetic algorithm, the selection and mutation operators are improved by introducing the simulated annealing algorithm, which enhances the global searching capabilities of the genetic algorithm.(4) With the feature selection methods proposed in this thesis for the analysis of experimental data, the optimal attribute set containing eight properties is obtained. Furthermore, a regression forecasting model is presented for the sensitivity analysis by using aforementioned eight properties data.更多还原