【作者】 侯澍旻；

【导师】 李友荣；

【作者基本信息】 武汉科技大学 ， 机械设计及理论， 2006， 博士

【摘要】 数据挖掘是一个复杂的过程，而针对时间序列具有维数高和动态性等特点，如何高效地对时间序列数据进行挖掘是一个十分有意义的研究课题。本文对时间序列数据非线性检验、降噪、分割等数据挖掘中的预处理工作进行了研究，并将时间序列数据挖掘技术引入机械故障诊断领域；通过应用实例，验证了这些方法和理论。论文主要完成工作如下：1)时间序列数据非线性检验研究提出了一种随机迭代修正幅值的傅立叶变换(SIAFFT)算法和KS检验统计量相结合的时间序列数据非线性检验方法，通过对弱非线性信号、强非线性信号以及含噪信号的检验，该方法均能得到正确的判断结果。与传统方法相比，该方法收敛速度快，具有较强的抗噪声能力，对非线性信号具有较高的敏感性。2)全局投影算法用于降噪和故障特征提取全局投影降噪算法具有良好的适用性，运算效率和降噪效果明显优于局部投影降噪算法。将全局投影算法应用于转子轴心轨迹的提纯，比基于谐波小波等的轴心轨迹提纯方法运算效率高得多。将全局投影降噪算法和共振解调技术相结合，提出了一种适用于低速重载轴承故障诊断的方法。利用此方法对某炼钢厂转炉倾动机构悬挂齿轮箱耳轴轴承进行故障诊断，诊断结果与实际情况相符。3)基于GG聚类的时间序列数据分割方法针对传统时间序列分割算法的不足，提出了一种基于GO聚类在线数据分割算法。该算法具有自合并功能，能不借助相关领域专家的支持自动寻找最优的分割子集数，是一种有工程应用价值的数据在线分割方法。4)基于KS检验的时序数据分类挖掘系统提出了一种基于KS检验的时序数据分类挖掘方法，在此方法基础上，建立了故障分类系统。通过仿真试验和齿轮及轴承的故障诊断，说明该方法在数据样本含有一定噪声时也能正确判断故障类型。该方法算法简单、计算效率高、实用性强，在机械故障智能诊断领域具有广泛的应用前景和推广价值。5)改进型支持向量机(ES-SVM)用于时间序列数据趋势预测将传统的基于支持向量机的时间序列预测方法进行改进，采用进化策略法搜索惩罚因子C，ε不敏感损失函数和高斯核参数σ。Lorenz信号的数值仿真实验预测结果证明改进后的ES-SVM算法比原SVM算法的预测精度更高。另外，由于在工程实际中所拾取的信号不可避免地会含有噪声，因此，将全局投影算法和ES-SVM相结合的预测方法具有更大的工程应用价值。6)构建基于时序数据挖掘的远程监测与诊断系统作为时序数据挖掘应用的重要组成部分，本文以机械在线监测系统为对象，对数据仓库结构、数据类型进行了探讨；并针对机械信号的特点，建立了时间序列数据挖掘快速处理模型。在此基础上，结合企业对具体远程监测与诊断系统的要求，构建了基于时间序列数据挖掘的远程监测与诊断系统，并成功应用于某烧结厂抽风机。更多还原

【Abstract】 Data Mining is a complex processing. The properties of time series are high dimension and dynamic. Hence, how to mine time series data effectively is an important research subject. Firstly, this dissertation researches the preprocessing work of Time Series Data Mining (TSDM), which includes nonlinearity test, denoise, segmentation. Then, some novel TSDM techniques are introduced into mechanical fault diagnosis field. Finally, this dissertation verifies these TSDM methods and theories via the simulation data and practical examples. The main works can be summed up as following:1) Time series data nonlinearity test researchA nonlinearity test method is proposed, which is integrates a Stochastic Iterative Amplitude Adjusted Fourier Transform (SIAFFT) algorithm and KS test Statistic. Not only weak nonlinear signal and strong nonlinear signal but signal with additive noise, the presented method can attain accurate test results. Moreover, through employing traditional test methods, both results show that this method is superior to the traditional ones in stronger robustness to signal with additive noise and higher sensitive to nonlinear signal.2) The globe projective algorithm and its application in noise reduction and the fault character extraction.A globe projective algorithm is introduced. Its calculation speed and denoise effect is superior to local projective. Based on these advantages, this algorithm is successful applied in purifying the exhaust fan’s rotor axis orbit in certain sinter plant. Moreover, according to the fault characteristic of low-speed rolling bearing, this paper proposes a new method in bearing fault diagnosis via integrating globe projective algorithm and resonance demodulation technique. By the proposed method, the rolling bearing fault of a converter trunnion was detected. The diagnosed result is consistent with the fact.3) Time series data segmentation method based on Gath-Geva clusterThis dissertation introduces an online data segmentation method based on Gath-Geva cluster. This algorithm can merge clusters by itself. The most advantage is that auto-searching the optimal segmentation subset without the expert’s supports. Hence, it is an important online data segmentation technique in engineering field.4) The time series data classification mining system based on KS testThis dissertation presents a novel data classification mining method based on KS test. By applying it in simulation test and gear fault diagnosis, this method can diagnosis faults though the data samples hold some noise. After using a few bearing time-domain fault data samples to building multi-class KS test classifier system, it classified kinds of faults precisely in a short time. The result shows this method has good classification ability and efficiency. It can satisfy the requirement of intelligent diagnosis.5) The modified SVM and its application in time series predictionThe dissertation modifies the traditional SVM. Through applying evolution strategies algorithm, the optimal penalty factor C,ε-insensitive cost function and Gauss kernel parameterσ. The forecasting result of Lorenz signal verifies the prediction accuracy of modified SVM (ES-SVM) method is higher than the traditional SVM. In addition, the measured practical signal often mixes noise. Hence, the global projective method can be used to reduce noise in advance. Then, applying the proposed method, the better forecasting result will be gained.6) Constructing the remoted monitor and diagnosis system based on TSDMAs a main part of TSDM application, this dissertation discusses data warehouse structure and data type of mechanical online monitor system. Then, the TSDM processing model is built. Based on it, according to the acquirement of a certain sinter plant, the remoted monitor and diagnosis system based on TSDM is constructed.更多还原