【作者】 孙东；

【导师】 施文康；

【作者基本信息】 上海交通大学 ， 精密仪器与机械， 2009， 博士

【摘要】 赤潮我国沿海最主要的生态灾害之一,随着经济的发展,环境污染日益加剧,赤潮爆发的频率逐年升高,赤潮的发生对沿海的生态环境和水产养殖造成了严重的影响。赤潮灾害的监测和预报作为一个国际性的难题已引起人们的高度关注和重视。为了对赤潮进行研究和预报,许多沿海国家都组建了自己的赤潮监测系统,由于赤潮过程的复杂性,人们目前还没有完全把握赤潮的发生、发展和消亡的相关机理,因此在众多现行的赤潮监测体系中,为了有效的处理大量的赤潮监测数据,机器学习和模糊技术被广泛使用。同时,为了更加全面的监控赤潮过程,多手段、全方位的立体监测系统的开发也势在必行,因而处理多源赤潮监测信息的不确定性并对其进行融合也将是一个难题。本文的研究得到国家高技术发展计划(863计划)赤潮重点监控区监控预警系统重大专项课题(2005AA635200)的资助,本文针对该赤潮监控系统获取得大量监测数据,结合机器学习、模糊数学、数据融合方法等理论,对赤潮监测数据的处理与分析方法进行了研究,主要研究内容和创新之处如下:首先,介绍了本项研究所依托课题的实施情况与五个监测子系统的实现技术,给出了多源监测数据的获取手段和传输方式。对于数据的预处理主要包括两项工作:一个是综合利用现有的相关标准对浮标传感器进行在线评价,并对符合线性分析标准的监测参数进行不确定度评估;另一个是针对监测数据中出现的缺失数据的情况,提出了一种适用于赤潮监测数据的时间序列的插补方法,取得了比一般插补方法更好的插补效果。经过对浮标传感器在线评价,发现一些在线监测参数满足与离线数据的线性回归分析标准,而另一些则不满足。满足标准的主要为温度和盐度数据,能够利用现场校正将其误差控制在允许范围内,因此,温度和盐度参数时间序列中的值可以用精确值表达,对于精确值时间序列,提出一种结合奇异谱分析和RBF神将网络的时间序列多步预测方法,取得了理想的预测效果。对于无法现场标定的参数(主要是叶绿素),需要考虑监测数据的不确定性,本文采用模糊时间序列方法对叶绿素时间序列进行预测。结合现有的二元模糊关系法和二元高阶模糊推理法的特点和优势,提出了一种改进的二元模糊时间序列预测方法,与其他方法相比,能够获得更高的预测精度。为了达到对赤潮状况进行快速预警的目的,利用本监测系统中获取的叶绿素多源监测信息,结合数据融合理论,本文尝试性的提出了应用于赤潮快速预警的多源叶绿素监测信息的数据融合模型,介绍了证据理论的主要内容,讨论了利用证据理论进行数据融合数据融和过程中的遇到的主要问题和现有的解决方法,结合赤潮预警的具体需求,提出了实用性的生成BBA和冲突处理的方法,做到了对赤潮发生的实时判断。赤潮生物密度与种群分布是赤潮监测中获取的重要数据,利用SOM神经网络对赤潮种群信息进行聚类,并将结果可视化,为专家分析提供了便利,讨论了fuzzy-ARTMAP神经网络算法,并将其与SOM神经网络相结合,利用其对种群变动进行了辅助预测,取得了良好的结果。考虑到赤潮专家分布于不同的城市,并且专家意见存在差异性,为了合理的表达和融合多专家意见,本文利用推广模糊数来表示专家的语言信息,给出了多专家意见的信息融合方法。针对现有模糊数相似性测度的缺陷,提出了一种新的专家意见之间的相似性测度的表示方法,并将其应用到多专家意见的信息融合方法中,利用数值算例论述了本方法的可行性。更多还原

【Abstract】 Red Tide is one of major ecological disasters of china coast. The monitoring and prediction of red tide is an open issue which has attract high attention. The related mechanism of red tide dynamic process is unclear for the complexity of red tide process, so machine learning and fuzzy sets theory is widely used in the current red tide monitoring systems. On other hand, in order to monitor red tide process more comprehensively, multi-method and stereoscopic monitoring system should be constructed, therefore handling and fusion multi-source monitoring information should be considered.The research work in this thesis is supported by“863”of high technology research and development program“monitoring and early-warning system for red tide key monitoring sea area”important specific subject. A great deal of data is obtained from the monitoring system, so combined with the theory of machine learning, fuzzy mathematics and data fusion, our research is emphasis on the handling and analysis the monitoring data. The major research work and contribution are as follows:First of all, the implementation situation of the subject and its five subsystems on which our research work based is introduced and the means to acquire and transmission of the data are described. The content for preprocessing of data comprise from two parts. In first part the buoy sensors are online accessed in which available related standards are comprehensive utilized, and the uncertainty evaluation is carried out for the monitoring parameters accordance with the linear analysis standard. In second part, in order to impute the missing data in monitoring time series, a new imputation method is proposed that can achieve better imputation result than common method.Being accessed to the buoy sensors, it is found that some on-line monitoring parameters accordance with the linear analysis standard, while some ones not. For the first kind, accurate value can be used to represent the data in temperature and salinity time series. A new time series multi-step prediction method combining singular spectrum analysis and radial basis function neural networks is proposed, and an exact result is achieved. For the second kind, considering the uncertainty of the data, fuzzy value can be used to represent the data in chlorophyll time series. Taking advantage of the superiority of two-factor time-variant fuzzy relation method and two-factor high-order fuzzy inference method, a better result is achieved comparing with other method.For the purpose of the early-warning of red tide, utilizing the multi-source monitoring chlorophyll information from this monitoring system, combined with the data fusion theory, a data fusion model which applied to the early-warning of red tide is tried. The major content of evidence theory is introduced, and the main problems and the existing solutions met in evidence theory application is discussed. According to the specific requirements, a new practical model for BBA generation and conflicts handling is constructed, which could realize the real-time judgment of red tide occurrence.The biological density and species distribution of the red tide are the important data derived from the red tide monitoring process. The SOM neural network is used to cluster the red tide population information and visualize the results which could facilitate the expert to analysis. The structure of fuzzy-ARTMAP neural network is talked about. The hybrid model of the fuzzy-ARTMAP neural network and SOM neural network is applied to aided prediction analysis the algae population fluctuation that have obtained good results.For the case that red tide experts are distributed in different cities with deferent opinion, extended fuzzy number is implied to express the linguistic information of experts, and multi-expert information fusion model is given. According to the defects of several existing fuzzy number similarity measurement, a new similarity measurement is put forward which has been applied in the multi-experts information fusion model, a numerical example is used to illustrate the efficiency of the proposed method.更多还原