【作者】 向秀桥；

【导师】 周建中；

【作者基本信息】 华中科技大学 ， 系统分析与集成， 2009， 博士

【摘要】 现代水力发电机组正向着大型、复杂、超临界参数方向发展，其运行的安全性、可靠性和稳定性已成为水力发电行业极为关注的研究热点和工程应用难题。随着人工智能理论研究的逐步深入和信息科学的蓬勃发展，国内外学者已在模式识别、信号处理方面获得了完备的理论和方法体系，这为故障智能诊断研究奠定了坚实的理论基础，同时，也对传统诊断方法提出了新的挑战。本文针对水电机组轴心轨迹识别、机组故障信号检测和提取分析中的科学现象与难点问题，引入混沌、支持向量机、粗糙集、信息融合等先进智能算法，深入、系统地开展了水电机组故障诊断研究，主要内容和创新性成果包括：(1)轴心轨迹作为水电机组状态的一种映射方式，其图形特征所包涵的信息反映了机组的运行状态。为提取轴心轨迹平移、旋转、尺度不变性特征，本文挖掘出Haar正交矩阵在图形识别中的独特优点，引入Haar正交矩阵分别对轴心轨迹的横、纵坐标进行快速变换，并利用变换后的对应系数得到轴心轨迹不同位置不同分辨率下的斜率，继而推求出相邻斜率直线之间的多分辨率夹角，所获夹角用于轴心轨迹的分层识别，其识别过程符合人类识别事物层次分析的规律。(2)考虑到轴心轨迹实时识别的要求及轨迹样本先验知识不足的工程应用现实，通过计算和变换距离向量，快速提取了轴心轨迹的Walsh谱特征，同时引入支持向量机理论学习谱特征和轨迹类型之间的映射关系，从而将训练好的分类器用于轴心轨迹识别。实例分析显示：该方法优于传统提取方法，具有较强的泛化推广能力和小样本学习能力。(3)为解决轴心轨迹识别中单一特征容易遗漏重要信息而复合特征会产生冗余信息的矛盾，提取、融合轴心轨迹的傅立叶谱特征和几何特征，并引入粗糙集理论对所得融合特征向量进行约简，推理出用于轨迹识别的规则体系。理论分析和实践结果表明，该方法充分体现了粗糙集的数据约简功能，能快速、准确地识别轴心轨迹。(4)Haar类正交变换(HTOT)算法高效，应用简捷，然而，HTOT数学方法的抽象性使其潜在的工程应用价值很少受到人们关注。为此，本文将HTOT理论应用于水电机组故障信号分析，并根据相应评价准则比较了HTOT与其它变换方法的应用效果，得出了不同快速变换适用于分析不同类型信号的结论，据此提出了根据故障信号类型选择变换的自适应诊断策略。在此基础上，采用粗糙集理论对变换系数组成的谱统计特征向量进行处理。研究结果表明，该方法简单易行，具有很强的针对性和灵活性，增强了故障诊断效果。(5)针对传统方法难以全面检测机组故障信号的不足，探索了非线性系统对初值敏感性的响应特性，建立了基于混沌振子与信息融合技术的故障信号检测方法：①采用混沌振子阵列确定出故障微弱信号的频率和相位；②引入多个不同的混沌振子和多种方法检测同一微弱信号的幅值，并对其检测结果加权平均融合；③综合运用统计距、二维熵和Walsh变换等方法从时域或频域角度对混沌振子的状态进行模式识别，并对其识别结果进行信息融合；④研究了变换系数的模极大值原理，提取了强幅值信号的奇异特征，并结合信号时域波形统计特征和混沌振子检测故障微弱信号的优点，提出了基于k／l技术的融合诊断策略，该策略更能全面反映水电机组信息，提高了故障诊断的准确度。更多还原

【Abstract】 The modern hydro-electric power unit is evolving toward the large, complex and supercritical parameter direction, its operation with safety, reliability and stability has become a research hotspot of great concern and engineering application problem in the hydro-electric power industry. With the in-depth study of the artificial intelligence theory and the rapid development of information science, domestic and foreign scholars have acquired a complete theoretical and methodological system in aspects of pattern recognition and signal processing, which has not only laid a solid theoretical foundation for the intelligent fault diagnosis technology, but also brought new challenges to the traditional diagnostic methods.For the scientific phenomena and difficult problems in the fault diagnosis of hydroelectric generating units, including shaft orbit recognition, signal detection and extraction, the advanced intelligent algorithms, such as chaos, support vector machines, rough sets, information fusion, etc, are introduced in this paper, and the in-depth and systematic diagnosis research has been performed, and the main contents and innovative results are listed as follows.(1) Shaft orbit, as a mapping way of status information, reflects the operation state of the hydroelectric generating unit from the viewpoint of its graphical features. To extract shaft orbit’s features invariant to rotation, scaling and translation, the unique strengths of Haar orthogonal matrix type are found, the horizontal and vertical coordinates of shaft orbit are fast proceeded by Haar transform, respectively, and the slopes with different resolutions are obtained at different positions by using the corresponding transform coefficients, and then the multi-resolution angles between the adjacent slope straight line are extracted skillfully and used for the hierarchical recognition, the recognition process accords with the principle of human’s recognizing things by using hierarchy analysis approach.(2) Taking into account the requirements of the real-time identification of the orbit sample and the engineering reality of the insufficient prior knowledge, the Walsh spectrum features of the shaft orbit are fast extracted by calculating and transforming the distance vector, then the mapping relations between the orbit features and the orbit types are learned by utilizing support vector machine theory, and the classifier gotten is used for the identification of shaft orbits. Examples analysis shows that the proposed method is superior to the traditional extraction methods, and has a strong generalization ability and small-sample learning ability.(3) To address the contradiction that it is easy to omit important information for a single feature extraction method while for the complex methods the redundancy information is often added, the geometric features and frequency domain features of shaft orbit are extracted and integrated, then the combination feature vectors are discretized, reduced and reasoned by introducing rough set theory, the rules achieved are used for the diagnostic test. The theoretical analysis as well as the experimental results fully displays the prominent data reduction function of the rough set theory, and shows the rapidness and accuracy of the proposed method.(4) Haar type orthogonal transforms (HTOTs), owning the efficient algorithm and the potential value in the engineering application, may be obtained conveniently. However, these advantages cause little attention of scholars’ since the mathematical abstract of HTOT. To this end, HTOT theory is introduced into the fault signal analysis, and the classification performance of its transform coefficients is compared with that of other transforms under the related evaluation criteria. As a result, the conclusion that certain transform fits the analysis of corresponding signals is drawn, and the adaptive diagnosis strategy that the transform has been adopted in accordance with the type of fault signal is offered. On this basis, the statistical feature vectors constituted by the transform coefficients are proceeded by rough set theory, the diagnostic test demonstrates that the proposed approach is flexible and effective, and has a very clear superiority over the traditional methods in both the diagnosis accuracy and speed.(5) Considering the defect that the weak signal cannot be detected comprehensively by the traditional methods, a fault signal detection method based on the information fusion and chaotic oscillator is put forward by exploring the response characteristics of the nonlinear system sensitive to the initial value.①the frequency and phrase of the weak signal are determined by the chaotic oscillator array.②various chaotic oscillators and various methods are introduced to the amplitude detection of the same weak periodic signal, then the detection outcomes are fused by the adaptive weighted fusion method.③ by using the statistics distance, two dimensional entropy and Walsh transform, the state recognition of chaotic oscillator is carried out from the viewpoint of time domain and frequency domain. Then the recognition outcomes are fused.④the singular features of the strong amplitude signal are extracted according to the modulus maximum principle of the transform coefficients, then combined with the signal time-domain statistical features and the advantage that the weak signal can be detected by the chaotic oscillator, the diagnosis method based on the k/l fusion technology is provided, which can represent information more comprehensively and diagnose fault more accurately.更多还原