基于流形学习的数据降维技术及工程应用研究

发布时间：2018-08-18 14:55

【摘要】：随着人工智能和大数据技术的飞速突破,机器学习和数据挖掘这些人工智能核心研究领域往往会得到普遍呈现出高维、非线性特性的数据。以模拟电路故障诊断为例,特别是在一个高度集成的电路板上进行板级故障定位时,由于元器件数量大且存在容差,导致采集到的数据更倾向于以高维数、非线性结构分布。对于这些大规模、高维非线性的数据,人们想要直观的感知隐藏在高维数据下的有用的知识,难度可想而知。数据降维,是高维数据维数降低的最有效途径之一。数据降维技术可分为线性降维和非线性降维。线性降维技术应用广泛,但对包含大量非线性数据的实际工程应用降维效果较差,这也使非线性降维技术逐渐成为当前的研究热点。为获取高维且难以理解的数据的低维表示,基于流形学习的局部线性嵌入维数约减方法利用局部线性和全局非线性的假设,使其在减少高维数据维数的过程中,仍可保持高维数据的原有结构。这种特性使其成为机器学习研究领域的研究热点之一。本文研究使用基于流形学习方法的局部线性嵌入(Local Linear Embedding,LLE)技术进行特征降维和特征提取的问题。针对模拟电路故障诊断工程中特征维数高问题,提出了一种基于小波包分解和LLE算法相结合的特征维数约简方案,以两级四运放低通滤波器电路为研究对象,通过基于克隆选择算法的故障诊断研究了故障特征维数约简技术。实验研究结果验证了本文提出的算法在模拟电路故障诊断特征降维问题的适用性,从而为LLE算法在复杂数据降维的工程应用提供了有用的参考。
[Abstract]:With the rapid breakthrough of artificial intelligence and big data technology, machine learning and data mining, which are the core research fields of artificial intelligence, often get high dimensional, nonlinear data. Taking analog circuit fault diagnosis as an example, especially in a highly integrated circuit board fault location, because of the large number of components and tolerance, the collected data tend to be distributed in high dimension and nonlinear structure. For these large-scale, high-dimensional nonlinear data, people want to intuitively perceive the useful knowledge hidden in high-dimensional data, it is difficult to imagine. Data dimensionality reduction is one of the most effective ways to reduce the dimensionality of high dimensional data. Data dimensionality reduction can be divided into linear reduction and nonlinear dimensionality reduction. Linear dimensionality reduction technology is widely used, but it has poor effect on practical engineering applications containing a large amount of nonlinear data, which makes nonlinear dimensionality reduction technology become a hot research topic at present. In order to obtain the low-dimensional representation of high-dimensional and incomprehensible data, the locally linear embedding dimension reduction method based on manifold learning makes use of the assumptions of local linearity and global nonlinearity to reduce the dimensionality of high-dimensional data. The original structure of high-dimensional data can still be maintained. This characteristic makes it one of the hotspots in the field of machine learning. In this paper, the problem of feature reduction and feature extraction using locally linear embedded (Local Linear embedded LLE technique based on manifold learning method is studied. In order to solve the problem of high characteristic dimension in analog circuit fault diagnosis engineering, a new feature dimension reduction scheme based on wavelet packet decomposition and LLE algorithm is proposed. Fault feature dimension reduction technique based on clonal selection algorithm is studied in this paper. The experimental results verify the applicability of the proposed algorithm in the characteristic dimensionality reduction problem of analog circuit fault diagnosis, which provides a useful reference for the engineering application of LLE algorithm in complex data dimensionality reduction.
【学位授予单位】：北方工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.4

【参考文献】