【作者】 姬鹏；

【导师】 张承慧；

【作者基本信息】 山东大学 ， 控制理论与控制工程， 2009， 博士

【摘要】 众所周知,21世纪是海洋世纪,海洋包括高新技术引导下的海洋经济将成为国际竞争的主要领域。在我国,赤潮灾害发生范围已遍及所有沿海省市,每年对海产养殖业、滨海旅游业等海洋经济产业造成的直接经济损失高达数以十亿元计。因此,研究赤潮分析预测关键技术和开发先进的赤潮预警应用系统对促进我国海洋经济可持续发展具有重要的现实意义和战略意义。赤潮是一种由多种因素综合作用引发的生态异常现象,具有随机性、模糊性、突发性和非线性等特点,对其建模和预测是非常困难的。赤潮现象的这些特点使得赤潮灾害的预测预警已成为当今海洋科学技术领域最具挑战性的难题之一。传统的研究思路大多数是从赤潮现象的某一个特点出发构建算法模型,并设计开发原型系统。这些模型只是从某一个方面反映赤潮的特性,而没有全面系统地刻画赤潮的本质和各方面的特征,实际应用的效果并不是很理想。为了克服传统研究思路的不足,需要不断引入新的理论和方法,同时需要将多种不同技术方法综合集成,从整体的观点对赤潮进行系统和全面的分析研究。本论文以山东省科学发展计划(重点项目)“海洋环境在线监测及灾害智能预警系统的研制(2004GG2205108)”为背景,在总结吸收前人研究成果的基础上,探索以知识发现的理论和方法为框架,多学科交叉综合研究,考虑将小波分析、主元分析、模糊聚类分析和神经网络等知识发现技术综合应用于赤潮数据资料的分析处理上,并在课题组众多研究成果的基础上设计实现赤潮知识发现平台,通过该平台构建多种组合模型,以期能从更多视角认识和掌握赤潮发生的内在规律,进而在赤潮预测预警研究上取得一定的成果。全文的主要内容包括以下几个方面:(1)首先在分析传统小波阈值去噪方法的基础上,提出一种基于平移不变的小波指数阈值去噪方法,并通过理论分析和仿真实验对其去噪性能进行了深入地分析和探讨,结果表明该方法能够有效克服传统硬阈值和软阈值去噪方法的不足,较为明显地提升了去噪性能,具有一定的通用性。进而针对赤潮分析预测中使用的数据资料可能受到大量噪声影响的问题,应用该方法对其进行去噪处理,获得了较高品质的数据资料,能够更为真实和客观地反映赤潮现象的本质特征,从而为后续的赤潮分析预测等提供了可靠和有效的数据源。(2)其次,在分析研究常规模糊c均值聚类(Fuzzy C-meansClustering,FCM)算法不足的基础上,针对FCM算法在聚类中心的迭代计算过程中,各样本点对聚类中心的确定具有相同的权重,容易导致少数异常样本的存在也会对聚类中心的确定产生较大影响的问题,提出了一种基于相似关系的模糊加权FCM算法(Fuzzy Weighting FCM,FWFCM),根据各样本对聚类中心的不同影响程度,为每一个样本设置了一个特征权值,使得各个样本在聚类中心的迭代计算中起到不同的作用。进而针对目前大多数赤潮预测算法不能合理区划赤潮起始、发生、发展和消亡的演变过程,应用FWFCM算法,找出比较可靠的赤潮生态过程所处阶段的范围划分规则,不仅得到了较准确的和符合实际的类别划分,直观和真实的反映了赤潮生态过程的模糊性和过渡性等特征,而且在分类后仍极大地保留了样本数据集的原始信息,为进一步探求分类的内在原因和解释样本点之间彼此关联的程度提供了有效的途径。(3)然后,针对传统的单一预测模型不能全方面刻画赤潮生态现象的本质特征,存在预测精度低和预测结果稳定性差的不足,以及神经网络预测模型对预测结果缺乏合理性解释的问题,提出了融合主元分析、模糊聚类分析、小波分析和神经网络的赤潮组合预测模型,使用主元分析方法和模糊聚类算法先将赤潮生态过程分解为若干不同本质的子过程,然后再应用小波网络对赤潮藻类密度等进行预测预报。该模型结构在一定程度上反映了赤潮生态过程的内在规律,可以对模型的输出结果给予更合理的解释。同时,在该组合模型在模型效率和预测精度方面都较单一预测模型有所提升。(4)最后,针对传统赤潮预测预警软件中普遍存在的模型单一化、数据与算法过耦合,以及可扩展性和可重用性差等问题,在深入研究知识发现过程机理、系统结构和运行机制,以及软件架构模型的基础上,提出了一种基于组件技术的面向服务的赤潮知识发现平台的解决方案,以期为用户提供一个可以快速和灵活构建各种组合模型,并能够测试和实际运行的赤潮分析预测的研究与应用环境。总之,本文在知识发现技术体系框架下,综合运用多种理论和方法,多视角、多层次系统全面地研究了赤潮的分析预测。大量实例分析表明,本文的研究思路具有较强的可行性和实用性,深化、丰富和发展了赤潮分析预测的理论和方法,为赤潮的预测预警研究提供了新途径。更多还原

【Abstract】 It is well known that the 21^st century is called "marine century". Marine and high-tech-driven marine economy is increasingly becoming major field of international competition.Red tide has brought about many negative environmental and economic consequences around,along the coastal areas of China in recent years.The annual direct economic loss caused by red tide has amounted to much as 10 billion Yuan average. Therefore,research on key technology of analysis and forecast for red tide,and development of red tide early warning application systems has important practical and strategic significance on promoting sustainable development of marine economy in China.Red tide is an anomalous ecological phenomenon caused by various complex factors and is characterized by random,fuzziness,abruptness and nonlinearity,so it is difficult to build up red tide prediction model. All the features of red tide make the research on the red tide prediction one of the most challenging subjects in the frontier field of marine science and technology.To overcome the shortcomings of traditional ways,new theories and methods must be continuously introduced.At the same time,different methods must be integrated to research red tide systematically and comprehensively in a macro view.This thesis is written against the background of the project named Marine Environment Observing and Calamity Intelligent Early Warning System,which is supported by Science and Technology Development Plan （Key Project） of Shandong Province（Grant No.2004GG2205108）.On the basis of previous achievements,by taking theories and methods of knowledge discovery as a framework,and cross-subject integrative research in analysis and prediction of red tide,this thesis applies knowledge discovery techniques（such as wavelet transform,principal component analysis,fuzzy clustering and neutral network） to data analysis and processing of red tide.At the same time,a red tide knowledge discovery platform is designed and implemented,thanks to numerous research achievements of our team.Hopefully,through construction of composition models based on the platform,researchers can fully and correctly understand the nature and internal mechanism of red tide from multi points of view,and then fruit certain achievements on the prediction of red tide.The thesis consists of the major novel technical contributions as follows:Firstly,grounded on the analysis of the traditional threshold methods,a de-noising method of wavelet exponential threshold is put forward,based on translation invariance.It proves that this method has effectively overcome the shortcomings of hard shrinkage and soft shrinkage function, and greatly improved the performance of de-noising.In response to the fact that the measurement data of red tide may be inevitably affected by noise and therein,the proposed method is applied to de-noise the data in order to achieve high-quality ones.Thus,the nature of red tide is more objectively and truly reflected through the de-noised measurement data, which provides a reliable and valid data source for the subsequent analysis and prediction of red tide.Secondly,this thesis studies fuzzy clustering theory and algorithms.In conventional FCM algorithm,every sample has the same influence on data set classification,but it is not often correct in practical classification process.This sometimes causes the fact that the ascertainment of cluster centers is strongly influenced by a few anomalous samples in data set.In order to solve the problem,a new fuzzy weighting clustering algorithm based on the fuzzy C-means algorithm and similar relation is proposed. When the fuzzy weighting coefficient is applied to the cluster centers and Euclidean distance,each sample has various influences on data set classification in FCM algorithm.The convenient prediction algorithm for red tide is not able to make reasonable distinction between the red tide evolutionary phases including its origin,occurrence,maintenance and dying.Therefore,the FWFCM algorithm is applied to find out the reliable division rules for different phases of evolutionary process of red tide. More accurate and authentic division is achieved,which reflects the fuzziness and transitivity of red tide intuitively and truly,and retains original information of the data sets.Thus,the proposed strategy provides a valid way to explore inherent reason of division and explain correlation degree between samples.Thirdly,a red tide combined prediction model is proposed,which is the integration of principle component analysis,fuzzy clustering algorithm, wavelet transform and neural network.The model reflects the inner mechanism of red tide to a certain extent,and explains the predictive result in a reasonable and clear manner.What’s more,this combined model efficiency increases the model efficiency and predictive accuracy.Fourthly,the inner mechanism of knowledge discovery system and software architecture model is studied.To solve the common problems （such as simplified model,coupling of data and algorithm,lack of expandability and reusability,etc.） of traditional red tide prediction software,a solution of red tide knowledge discovery platform is put forward,based on component technology and service oriented architecture.Through the convenient environment provided by the proposed solution,users could build various combined models,test models,and execute in actual business.In short,the thesis takes theories and methods of knowledge discovery as a framework,and integrates cross-subject research in analysis and prediction of red tide.Lots of case studies show that the research has enriched and deeply developed the theory and method of the analysis and prediction of red tide with its practicability and effectiveness.It provides a new approach to the research on red tide prediction.更多还原