基于流形学习和极限学习机的柱塞泵复合故障诊断的研究

Along with the big scale, multi-functionality, complexity on hydraulic system, people begin to pay more attention to the reliability and safety of the system. As the power source of hydraulic system, piston pump is called ‘the heart’ of the system, and the quality of the hydraulic system is directly affected by the working condition of the piston pump. Many methods have been proposed in the literature on fault diagnosis of piston pump. However, some of the methods can distinguish only one type of fault based on vibration signal only; while some others can distinguish various faults when they occur individually. In real applications, due to the harsh working environment, multiple faults on piston pump are very likely to occur simultaneously. Hence, the diagnosis methods on multi-fault of piston pump are much more desired. Therefore, this project proposed a new diagnosis method based on manifold learning and extreme learning machine (ELM) for multi-fault diagnosis, which could achieve multi-field feature fusion and ELM ensemble decision fusion. The proposed method provides a unified framework for multisource heterogeneous information fusion and extracts features from multiple fields, such as time, frequency, and time-frequency field. In addition, we conduct the chaotic features analysis on vibration signals in order to extract the chaotic features. Further, a research on activation function of ELM network is conducted and a wavelet based function is introduced to play as the activation function. In all, this project may provide a new sight on multi-fault diagnosis of hydraulic pump.

随着液压系统的规模、功能、复杂程度及自动化水平的日益提高，对系统的可靠性和安全性的要求也越来越高。柱塞泵是一些大型高压液压系统的动力源和故障源，对柱塞泵的故障检测和诊断对高压液压系统可靠性的提高有重要的意义。由于柱塞泵的故障具有隐蔽性、多样性和因果关系复杂性等特点，柱塞泵复合故障的诊断更加困难。针对目前柱塞泵复合故障诊断中的信息采集、复合特征难于分离和复合故障中各故障识别精度不高的难题，本项目提出利用多传感器采集多源信息；利用多信息域的信号处理方法、混沌理论和流形学习算法提取能全面反映柱塞泵运行状态的高维故障特征集；利用非线性流形学习的方法对故障特征进行维数约简，同时使得不同类样本特征在新特征空间易于分离；利用极限学习机集合在决策层融合的方法，充分利用各传感器之间的互补信息，对柱塞泵的复合故障进行识别。本项目的实施将为机械故障中复合故障的诊断提供一定的理论基础和技术支持。

柱塞泵是液压系统的核心部件，它能否正常稳定的工作会直接影响整个液压系统的运行状况。由于柱塞泵运行状态复杂、工作环境恶劣，其状态信号具有非线性程度高、噪声干扰强等特性，其复合故障状态难以识别，致使目前柱塞泵故障诊断的理论和方法还远满足不了实用化的要求。本项目通过对柱塞泵常见故障模式的故障机理分析，了解各故障的演化和征兆。项目提出了基于数据驱动的柱塞泵健康状态评估以及故障诊断方法，该方法采集泵壳振动信号、泵出口压力信号和泵出口流量信号作为原始信号，对原始信号进行时域分析和时频域分析得到特征矩阵，通过局部切空间排列算法对特征矩阵进行约简，利用基于极限学习机算法的健康状态评估模型与故障诊断模型对柱塞泵进行健康状态评估、单故障诊断、多故障诊断包含复合故障诊断。项目将故障诊断方法应用于柱塞泵滑靴磨损故障的故障诊断、轴承故障的故障诊断以及柱塞泵多故障诊断。另外，项目中对极限学习机及其衍生算法进行了改进，提出了基于引力搜索算法的极限学习机模型选择方法和基于υ正则化的优化极限学习机。在柱塞泵故障诊断方法的研究中，针对非均衡数据集，提出了针对非均衡数据集的基于加权极限学习机的轴承故障诊断方法；针对故障诊断中部分样本无标签的问题，提出了基于半监督学习的故障诊断方法，采用了基于引力搜索算法的聚类方法给无标签样本加上标签。本项目提出的故障诊断方法，不仅适用于柱塞泵的故障诊断，还适用于其他机械设备的故障诊断；本项目提出的极限学习机及其衍生算法的改进方法，适用于与数据压缩、特征学习、聚类、回归和分类相关的多种应用。