风机齿轮箱振动信号的稀疏分解理论与其损伤程度定量评价研究

The condition monitoring and fault diagnosis of wind turbine gearboxes is very important for the stable operation of wind turbines. Aiming at the complex characteristics of their vibration signals, and basing on the synthesis of several subjects, such as information science, applied mathematics, mechanical science, et al, this project aims to put forward the sparse decomposition theory of the vibration signals and the quantitative evaluation method of damage degree for the wind turbine gearboxes. The specific research content is as follows: Establish the standard feature waveform library and the coupling characteristics waveform library respectively for the typical faults and the compound faults. Use the above waveform library as the matching target, the building principle of the dictionary library which consist of many types of basis function is proposed. Build the regularization model and determine its regularization parameter for the signal sparse decomposition, and then come up with the fast algorithm. Based on the amendment of the decomposition coefficients with its singular value, a noise suppression method is presented. And based on the signals' multidimensional feature vector, a quantitative evaluation method for the injury degree of the wind turbine gearboxes is put forward. The innovative achievements included the following: Based on the study of signal sparse decomposition, the diversity damage characteristics which may have consisted in the wind turbine gearbox vibration signals is enhanced, decoupled and separated. Furthermore, the theory and techniques for the realization of the damage pattern classification and the quantitative recognition for the damage severity are formed. This research project is of great value to the theory development and the engineering application, will effectively enhance the service performance of the wind turbine and avoid the economic loss caused by the failure downtime.

风机齿轮箱的状态监测与故障诊断对于机组的安全稳定运行具有重要意义，本项目针对其振动信号的复杂性特点，结合信息科学、应用数学、机械科学等多学科交叉，拟提出风机齿轮箱振动信号的稀疏分解理论与其损伤程度定量评价方法。具体研究内容： 建立风机齿轮箱典型故障标准特征波形库和多故障耦合特征波形库，并以此为匹配目标，提出与风机齿轮箱多样性损伤特征相匹配的多类基函数字典库构建原理；建立信号稀疏分解的正则化模型并确定正则化参数，研究模型的快速求解算法，提出基于信号稀疏分解系数奇异值修正的噪声抑制办法；提出基于信号多维特征向量的损伤程度定量评价方法。创新成果：基于信号稀疏分解的风机齿轮箱多样性损伤特征增强、解耦与分离技术；风机齿轮箱损伤程度的定量识别理论与技术。研究成果可为风机齿轮箱故障预示与运行安全评估提供有效的分析手段，避免因机组紧急停机而带来的损失，具有重要理论意义与工程实用价值。

本项目针对传统的基于频带滤波的共振解调技术存在的局限性问题，提出基于多分辨率奇异值分解（MRSVD）的包络解调方法，并将其应用于强背景噪声及谐波干扰影响下的轴承振动信号的解调分析。与传统的基于频带滤波的共振解调方法相比，该方法无需预先知道故障引起的共振频带信息，可避免噪声、谐波振动、周期性冲击振动的频率不可分离问题，且该方法计算效率高，除了计算行数为2的矩阵奇异值分解外，该方法仅涉及矩阵的加、减运算。该方法在轴承早期微弱故障的冲击特征检测中具有很好的应用前景。.线调频小波路径追踪（MSCPP）方法可从非平稳振动信号中精确提取瞬时频率，但该方法由于盲时频全局搜寻最佳线调频小波原子（Chirplet），致其计算效率低。本项目将同步挤压小波变换（SST）和MSCPP方法联合应用于转速波动条件下拾取的轴承非平稳振动信号的解调分析。基于SST的时频谱分析可为MSCPP方法参数的合理设置提供依据，避免MSCPP方法盲全局计算造成的时间浪费。该方法可扩展应用于不方便安装转速计时，转速波动状态下的旋转机械故障诊断。.本项目提出一种提高深度学习诊断模型识别效果的前期数据预处理方法－基于频谱包络曲线的稀疏自编码算法。该方法对采集到的时域信号经频域处理后，预先对低层频域信号提取包络线，得到表征频域变化态势的信息成分，接着再与稀疏自编码结合构建稀疏自编码的故障诊断模型。齿轮箱故障诊断实验证明，与原始频域输入相比，所提方法能够在保证诊断效果的同时，降低计算复杂度和所需要的存储空间。.随机森林（RF）算法的运行效率和识别精度受决策树数目的影响, 本项目提出了一种改进的RF算法-极端随机森林(ERF)法，该方法通过随机映射矩阵来随机降低高维数据的维数。从分类的角度看，该方法通过降维来提高分类性能，大大减小了输入随机森林的样本维数，大大提高了RF方法的计算效率，同时伴随的特征增强过程保证了随机森林的识别效果。本项目将ERF方法应用于3D打印机和风力发电齿轮箱的故障诊断实验，验证了ERF方法的性能。