基于多尺度leaders多重分形与多尺度约束PCA的汽车起重机主泵特征提取方法研究

Truck cranes are widely used in numerous infrastructure projects. Because of the severe working condition, truck crane is one type of special mechanical equipments with a highest accident rate. Plunger pump is a key assembly of truck crane, and promptly and accurately dealing with the equipment breakdown is very important in terms of reliability and downtime decreasing. Extracting the features relevant to the equipment conditions from mechanical signals including abundant running information is crucial to fault diagnosis. This project takes the vibration signal of the plunger pump in truck crane as study object. Because the signal has a typical non-linear and non-stationary character, a novel multifractal features extraction method based on multiscale decomposition leaders is proposed. We begin with the wavelet leaders, and then the research is deeply into the frequency band adaptive decomposition, including empirical mode decomposition (EMD) and wavelet packet decomposition, respectively. Based on the capacity of intensive depiction of signals' local characteristics, and the fault diagnosis prior knowledge of mechanical equipment, a multiscale constrained principal component analysis (MCPCA) based conception of features dimensionality reduction is proposed, and a corresponding unified projection frame is established. Furthermore, the MCPCA-based features dimensionality reduction algorithms based on Singular Value Decomposition (SVD) and Machine Learning Model (MLM) are developed, respectively. The redundant information in the feature space is removed effectively, and the constructive information is stand out. This research might provide theoretical proposals and technology supports for improving the precision of fault diagnosis for plunger pump in truck crane industry.

汽车起重机广泛应用于基础建设项目，工作环境恶劣，是事故率最高的特种机械设备之一。柱塞泵是起重机液压系统的关键部件，及时、准确地对其故障进行诊断是设备安全可靠运行的重要保障，而提取有效反映设备运行状态的特征是准确诊断的关键。本课题以汽车起重机主泵振动信号为研究对象，鉴于其典型的非平稳、非线性特性，提出基于多尺度分解系数leaders的多重分形特征提取方法。从基本的小波leaders出发，进一步提出频带自适应的经验模式分解和小波包分解系数leaders的多重分形特征提取模型。在利用多尺度分析刻画信号局部特性的同时，结合机械设备故障信息的先验知识，提出多尺度带约束主成分分析特征降维概念，建立对应的特征空间投影框架，开发基于奇异值分解和机器学习算法的多尺度带约束主成分分析特征降维算法，消除特征中的冗余信息，突出有效特征。为提高汽车起重机主泵故障诊断的水平提供进一步的理论和技术支持。

汽车起重机广泛应用于基础建设项目，工作环境恶劣，是事故率最高的特种机械设备之一。柱塞泵是起重机液压系统的关键部件，及时、准确地对其故障进行诊断是设备安全可靠运行的重要保障。本课题以汽车起重机主泵为研究对象，对其多尺度分形特征提取方法及故障诊断理论进行了研究。在多尺度分形特征提取方面，鉴于振动信号典型的非平稳、非线性特性，从基本的小波leaders出发，提出了基于小波leaders的多重分形振动信号特征提取及表示方法。基于bootstrap技术的最优块长求解算法，建立了振动信号小波leaders多重分形特征的统计性能分析方法。研究结果表明, 小波leaders多重分形特征能够很好的反映振动信号的几何结构特征, 基于块bootstrap方法能有效分析多重分形特征统计性能。进一步利用经验模式分解提取带噪信号的多尺度趋势，获取信号的波动特征，进而得到信号的多尺度分形特征。利用双树复小波变换的近似平移不变性、有限的数据冗余性、完全重构性和计算效率高等良好特性，提出基于双树复小波的多重分形特征提取方法，与传统的多项式拟合去趋势多重分形方法相比，自适应趋势提取方法能根据信号自身特点确定信号的趋势和不重叠等长度子区间长度，所得结果更加精确。在特征降维方面，结合故障先验知识，运用基于距离评估技术的特征选择方法，提高故障诊断的精度。针对增量式属性约减问题，提出了基于二进制分辨矩阵的属性约减算法，与同类算法比较表明，该算法不仅可以快速得到约减结果，而且具有较小的空间复杂度，并且能够处理实验样本逐渐增多的情况。对柱塞泵等旋转机械提出了适用于在线实时诊断的模型，包括基于粒子群优化的相关向量机（PSO-RVM）方法和基于反向跟踪算法的神经网络方法。提出了适用于异常检测的SVDD-聚类方法，结合粒子群算法对SVDD模型参数进行优化，对检测到的异常数据，利用基于DBI指数的聚类算法，对故障类型进行无监督机器学习方法，自动识别出故障类型。