【作者】 杨勃；

【导师】 卜英勇；

【作者基本信息】 中南大学 ， 机械电子工程， 2010， 博士

【摘要】 陆地资源日趋枯竭,大洋钴结壳作为一种极具商业开采价值的海生矿产资源,获得了美国、俄罗斯、德国、日本等西方发达国家的重视。随着钴结壳勘探开采技术研究逐渐深入,部分国家已经步入试开采阶段。与西方发达国家相比,我国勘探开采研究工作起步较晚。当前,国际社会对海洋资源争夺日益激烈,为维护我国海洋权益,开辟我国新的矿产资源来源,有必要研究钴结壳的探测采集技术与装备。钴结壳矿床底质自动识别是实现钻结壳高效开采的一个重要环节。本文在国家自然科学基金项目“深海钻结壳微地形监测技术与最佳采集深度建模研究”的资助下,对基于超声探测手段的深海钴结壳矿床底质识别的相关技术进行研究。作者首先从海底沉积物声学探测分类技术入手,查阅大量相关文献,并根据采矿车车载探测的特点,确定采用本文介绍的超声探测软硬件实验系统通过正入射法对钴结壳矿区内23种底质进行识别。针对钴结壳矿床底质超声探测识别关键技术,本文进一步从特征提取,特征级融合以及非线性分类识别三个方面进行了深入的研究：1)钴结壳矿床底质回波特征提取技术研究作者首先将成功用于水下沉积物有代表性的基于时频域特征提取方法：正交小波域时间子带能量特征,尺度子带能量特征,平稳小波域奇异值分解特征,模极大值边缘特征尝试用于钴结壳矿区底质回波特征提取。此外,基于信号复杂性熵度量方法,本文还引入了一种平稳小波域多分辨率奇异谱熵特征提取方法。实验结果表明,对于表面地形起伏的部分底质,上述特征出现了不同程度的退化。进一步,为消除钴结壳矿床部分底质表面地形起伏因素对特征稳定性的影响,基于信号稀疏分解技术,提出了一种样本字典域类别能量特征提取方法。在比较实验中,针对钴结壳矿床底质,类别能量特征取得了最好的识别效果。2)特征级融合方法研究及其在钴结壳矿床底质识别中的应用进一步,作者对特征级融合理论与技术进行了深入的研究,尝试利用已有的小波域特征提取方法,采用特征级融合方法提高钴结壳矿床底质的识别效果。首先介绍了三类有代表性的特征级融合方法：串行融合,并行融合和矩阵融合方法。然后,从fisher准则的角度分析了三种特征级融合方法之间的关系,发现后两种特征融合方法本质上都是特定约束条件下的串行融合方法,并指出除了串行融合方法具有融合的fisher稳定性之外,其它两种方法均不具备fisher稳定性,因此有可能出现退化。此外,从bayes最优性条件的角度对三种特征级融合方法进行了研究,指出后两种特征级融合方法要达到bayes最优所需满足的条件要比串行融合方法要更为严格。最后,在串行融合框架下,提出了一类快速串行融合方法,该方法具有fisher稳定性,且无需更为严格的bayes最优性条件。钴结壳矿床底质识别实验表明,采用串行框架下的特征级融合方法在识别效果上有较好的表现。3)基于核空间的非线性分类识别研究及其在钴结壳矿床底质识别中的应用最后,作者对基于核空间的非线性分类识别技术进行了研究,并尝试从优化分类边界和多核信息融合的角度实现更优的钻结壳矿床底质识别效果。首先,本文对高斯核参数学习进行了研究,提出了两种高斯核参数学习准则。为进一步提高非线性分类识别效果,作者从引入局部信息和多核信息融合两个方面对基于核空间的非线性分类识别技术进行研究。在现有的流形正则化最小二乘机基础上,提出了一种改进的凸优化流形正则化最小二乘机,并给出了对应的优化算法和其核化版本。对多核学习方法,作者尝试在核采样空间对多核学习方法进行研究,提出了有代表性的三种多核学习算法和基于FSM2准则的多核融合参速学习方法。钴结壳矿床底质识别实验结果表明,上述方法能够在一定程度上改善识别效果。更多还原

【Abstract】 Along with the shortage of land resources, as a kind of important ocean resources, cobalt-rich crusts(CRC) have been drawn much focus on by western advanced countries, such as America, Russia, German, Japan, etc. Based on the early related research achievements, Some countries have stepped into the trail stage of exploitation. By contrast, the related research work in our country is still in the initial stage. In order to protect our country’s ocean right and obtain CRC resource for our country, it is necessary to do research on the exploring and exploiting techniques and equipments.CRC recognition is an important link of realize the exploiting work highly effectively. Under the support of the project of the National Natural Science Foundation of China, "Research on detection technology of deep-ocean cobalt-rich crusts micro-terrain and best cutting-depth controlling model", the author tried to do research on the related techniques of underwater materials classification and recognition using sonar exploring method.The author started with underwater sediments classification. After consulting a large quantity of related research papers and technique reports, the author determined to complete the classification and recognition of 23 kinds of underwater materials using the sonar exploring experiment system. At the further step, the author tried to do research through the following three approaches.1) The techniques of feature abstraction of echoesThe author tried to use the representative sediments echo feature abstraction methods in time-frequency domain:time sub-band energy feature in orthogonal wavelets domain, scale sub-band energy feature in orthogonal wavelets domain, singular value decomposition feature in stationary wavelets domain, wavelets modulus edge to abstract the features of above mentioned 23 kinds of underwater materials. Besides, a new kind of echo feature abstraction method named multi-resolution singular spectrum entropy was introduced. At the further step, in order to remove the influence of the heavily uneven surface of underwater materials, a novel kind of echo feature abstraction method named class energy feature in echo samples dictionary domain based on the theory of signal sparse decomposition is proposed. Among all above echo features, using the class energy feature the best classification and recognition results were obtained in the related experiment.2) The techniques of feature level fusion and its application to echo recognition of CRC depositThe author tried to do research on the techniques of feature level fusion in order to improve the effects of classification and recognition. Firstly, three representative feature level fusion methods:serial fusion, parallel fusion, matrix fusion method are introduced. Then the relationship among these three fusion methods was analyzed using Fisher criteria and one conclusion was drawn that the last two fusion methods are both a kind of special serial fusion methods with special restriction. Besides, the author indicated that the last two fusion methods are not of fisher stability except serial fusion method, which means using the last two methods are possible to lead to the degenerating results. At the further step, the sufficient conditions when the three fusion methods would be Bayes optimal for two-class classification are researched. And the research results show that the stricter conditions are needed for the last two fusion methods than serial fusion method. At last, a kind of fast Serial fusion technique based on discriminant space which can ensure none degenerateness was proposed.3) The techniques of nonlinear classification and recognition in kernel space and its application to echo recognition of CRC depositThe author tried to do research on the techniques of nonlinear classification and recognition in kernel space. At first, two Gaussian kernel parameters learning criteria were proposed. Then the improving techniques of nonlinear classification and recognition in kernel space were considered by using local information and multiple kernel fusion method. Based on the proposed discriminatively regularized least-squares classifier, a modified discriminatively regularized least-squares classifier mode which leads to a convex optimality problem was proposed. At the further step, the author designed the related optimality algorithm and obtained the related model in kernel space. Besides, the author tried to do research on multiple kernel learning methods in kernel sampling space. A kind of fusion parameters learning methods using FSM2 criteria and three multiple kernel learning algorithms in kernel sampling space were proposed. The experimental results of classification and recognition of the underwater materials such as CRC etc show that all above improving techniques can improve the classification and recognition results in certain extent.更多还原