【作者】 吴绍春；

【导师】 吴耿锋；

【作者基本信息】 上海大学 ， 控制理论与控制工程， 2005， 博士

【摘要】 地震预报是一个国际公认的世界性难题。我国地震预报事业经过30多年的发展,积累了丰富的宝贵经验和大量的数据资料,全国的地震台网更是每日都在记录着数以千兆计的海量地震前兆观测数据。本文将数据挖掘技术引入到地震预报领域中,研究现有地震数据处理与数据挖掘技术交叉结合的方法,充分应用现代高性能计算环境,从这些海量数据中挖掘出地震预报所需的规律性知识,以便辅助领域专家提高地震预报的准确性。在探讨现阶段数据挖掘算法模型及其实现基础上,本文首先对地震预报的传统方法(地震震例数据和前兆观测数据分析)进行探讨。同时,围绕地震地区相关性分析、地震序列分析和地震前兆的规律性认识等关键问题进行分析研究,实现并行关联规则算法、基于地震相似度的时间序列相似性匹配算法以及序贯模式挖掘算法。然后,基于时序分析技术,提出一系列地震前兆观测数据处理模型和并行实现算法。最后,结合实际应用实现一个地震预报并行数据挖掘平台,为地震预报数据挖掘的海量数据处理提供强大技术支持。本文的主要创新性工作包括:1.基于关联分析技术研究地震相关地区的搜索方法,提出并实现了一种基于主从模式设计的并行关联规则算法FPM-LP( Fast Parallel Mining of Local Pruning)。本文把地震地区相关性问题转化为时间序列的关联规则挖掘问题,通过相关的实验和结果分析,挖掘出许多有价值的地震区域相关性知识。2.基于时间序列相似性匹配技术对地震地区相关性进行分析,实现了基于相似度的地震时间序列相似性匹配算法WSM3S (Whole Sequence Matching Based-on Seismo Similiarity Support)。本文从地震三要素时、空、强的三维角度,给出了地震相似度定义和时间序列相似性匹配模型及算法。通过分析近二十年来我国地震活动频繁区域的历史数据,应用该算法进行多种不同粒度、不同时间差的序列相似性实验分析,取得了可信度较高的结果。3.基于序贯模式挖掘技术进行地震序列分析的研究,提出并实现一种基于广义约束规则的序贯模式挖掘算法SPBGC(Sequential Pattern Mining Based on General Constrains)。本文将地震序列的相关领域知识定义为一组广义约束规则,应用该算法从地震震例数据中挖掘广义地震序列,为领域专家进行地震序列的相似性研究提供强有力的支持。4.基于时序分析技术重点研究地震前兆观测数据的处理方法,提出一系列实用地震前兆观测数据处理并行实现算法。首先,提出基于动态规划的时间扭曲方法进行子序列搜索的相似性度量,能有效地进行考虑噪声、幅度、偏移等问题更多还原

【Abstract】 Earthquake prediction is a worldwide challenging problem. With the development of earthquake prediction in the past 30 years, a large amount of prior knowledge and billions of data have been accumulated in our country. The gigantic auspice data under earthquake conditions is recorded by the sensor network of seismological observatory everyday. In this paper, we introduce the advanced data mining techniques into the earthquake prediction field, and several novel approaches between data mining and seismological data analysis are studied. Meanwhile, just by using the techniques of high performance computing and parallel data mining, seismological domain knowledge hidden in the gigantic data can be efficiently discovered to support earthquake prediction, therefore the accuracy of the earthquake prediction can be improved effectively.On the basis of discussing the existing data mining algorithms, the paper mainly focuses on the domain knowledge of seismology and the traditional methods for earthquake prediction. Then, by using the relativity analysis on earthquake zones, the earthquake sequences and the rules of earthquake auspice data, it carries out several parallel data mining algorithms such as association rules based parallel mining algorithm, the seismological similarity and similarity-matching algorithm realization, and the sequential pattern mining algorithm etc. Furthermore, the earthquake auspice data processing method and a series of parallel implement algorithms are proposed based on the technique of time series analysis. Finally, the parallel seismological data mining platform is implemented, which integrates all of the algorithms proposed in this paper.The main contribution of the dissertation is shown as follows:1. By analyzing and discovering the earthquake catalogue data on the relativity of earthquake zones, a Master/Slave mode based parallel mining algorithm FPM-LP (Fast Parallel Mining of Local Pruning) is put forward by using association rules, just as well as the relative preprocessing algorithm is presented. The experimental results demonstrate that the algorithm is satisfactory to find relative earthquake zones.2. On the basis of analyzing the relative earthquake zones on the technique of time series similarity matching, the seismological similarity and similarity-matching model on the relative earthquake zones and its algorithm WSM3S (Whole Sequence Matching Based-on Seismo Similarity Support) are proposed according to the three earthquake essential factors, which are named time, space, intensity separately. Just by更多还原