【作者】 孙欣；

【导师】 丁香乾；

【作者基本信息】 中国海洋大学 ， 信号与信息处理， 2006， 硕士

【摘要】 知识发现(Knowledge Discovery in Database,KDD),是一种以计算机为工具,将人工智能、统计、计算机及数据库等技术相结合,旨在从数据中提取总结出新信息的技术。感觉评估,即测量、分析和解释人感觉(视觉、味觉、嗅觉、触觉、听觉)对产品特征的反应或评价的科学技术。感觉评估具有主观性、灵活性的特点,以前通常依靠专家经验进行。随着信息量的膨胀、工业需求的增加和评估精度的提高,传统方法已经难以大规模推广。因此,借助神经网络、统计分析、数据挖掘等技术手段实现智能感觉评估中的知识发现,将具有重要的现实意义。神经网络具有非线性、自学习和推广能力强的特点,但经训练获得的知识不容易显式表达,难以被人们理解。为此,首先引入了神经网络函数逼近器的知识发现算法,将隐层单元的非线性激励函数进行三段线性化逼近,实现了输入与输出之间关系的规则提取。通过对输入神经元进行削减实现了输入空间的压缩。将该算法应用于卷烟感觉评估,提取了烟叶理化指标与感官质量指标之间的规则。但此方法仍有一些不足,为了使提取的规则和知识更加形象化、精确化,接下来引入了M5算法。在分析了该算法的原理、流程之后,将其应用在卷烟的感觉评估领域,实现了烟草数据知识发现。但在实际应用中,只采用一种方法往往难以达到理想的效果。最后,将统计分析、神经网络与M5模型树方法进行比较与融合,建立了烟草行业感觉评估的知识发现整体方案。该方案不仅通过了实验验证和实践证实,而且已经在企业中被采用,辅助专家进行感觉评估工作。更多还原

【Abstract】 KDD is a technology, which uses computer and combines artificial intelligence Statistics and computer with database and other technologies to extract new information from data. Sensory evaluation has subjective and agilitive characters, which once depended on experts’experiences. With information expanding, industry demands increasing and improve of evaluation precision, traditional methods have hardly been spreaded extensively. So using neural networks, statistics and data extraction to realize intelligent sensory evaluation for KDD will be important and significant.Neural networks have characteristics of nonlinear mapping, self-learning and generalization. However, the knowledge extracted from trained networks is hardly expressed in formulas. So it is difficult to understand. Firstly, neural networks pruning method can achieve inputs dimension compression and feature extraction. The activative function of the hidden unitis then approximated by a three-piece linear function. This method can apply for tobacco sensory evaluation to extract input indexes which are more correlative with output indexes and the piecewise linear rules. But this method has some deficiencies. In order to make the extracted rule and knowledge more visual and precise, M5 method is then introduced. After analyzing the principle and flow of the arithmetic, application in tobacco sensory evaluation for KDD is achieved. Finally, statistics, neural networks and M5 method are comparaed and combined to design a solution project of KDD for tobacco sensory evaluation. The project has not only been validated by experiences but also confirmed by practice. It has also been accepted by a company and helped exports apply for sensory evaluation.更多还原