变转速下旋转机械故障状态的稀疏表示与深度学习诊断方法研究

In the operation process of rotating machinery, such as aero-engine and gas turbine, the change of environment and operating state (fan inlet distortion, combustion chamber afterburning) or fault condition initiation (compressor rotor-stator rubbing, intermediate bearing spalling) will lead to speed variation in a large range and also small instantaneous speed fluctuation, the monitoring data of all components exhibit non-stationarity significantly, while the time-frequency characteristics become blurry, thus results in difficulty of monitoring and further increases risk of system deterioration. In order to perform fault diagnosis of rotating machinery under variable speed, a method of feature extraction and diagnosis based on time frequency analysis will be established. Firstly, appropriate basis functions will be selected incorporating the waveforms of key faulty components like planetary gear, rotor, bearing; secondly, based on time-frequency atoms, a sparse dictionary will be constructed through data training; and then the sparse coefficients will be obtained by reconstruction algorithms following ridge path, inverse transform will be performed to acquire sparse expression of fault state in time domain; finally the sparse samples of fault states will be utilized to establish the deep neural network classification model, after an optimization process of the structure parameters, the precise fault diagnosis of planetary gear - rotor - bearing system could be accomplished.

在航空发动机与重型燃气轮机等旋转机械的运行过程中，由于环境与工作状态的改变（如风扇进气畸变、燃烧室加力燃烧）或故障状态的出现（压气机动叶-机匣碰磨、中介轴承局部剥落）会导致其转速的大范围变化以及瞬时转速的小范围波动，进而使得各个部件的监测数据表现出显著的非平稳性，时频域特征模糊，提高了系统的监测难度与故障进一步恶化的风险。本课题针对旋转机械在变转速下的故障诊断问题，建立一种基于时频分析的特征提取与诊断方法。首先针对行星齿轮、转子、轴承等关键部件的故障时域波形选取适当的基函数；其次以基函数的时频变换形式作为备选原子，通过数据训练建立稀疏字典；然后以重构的时频域脊线为路径进行监测数据的稀疏特征提取，反变换得到故障状态的时域稀疏表示；最后将各故障状态的时域稀疏表示作为样本建立深度神经网络分类模型，优化模型结构参数，实现行星齿轮-转子-轴承系统的精确故障诊断。

航空发动机、燃气轮机以及风力发电机等能源与动力装备目前在我国国民经济发展进程中扮演着不可或缺的重要角色。随着对能源动力装备性能要求的进一步提升，保证机组的长期安全稳定运行成为了一项至关重要的工作。本项目针对典型旋转机械的行星齿轮-转子-轴承系统在复杂工作环境中存在瞬时变化工况的特点，重点分析瞬时转速变化对转子系统振动响应的影响，建立适用于瞬时状态特征提取的监测数据时频表示，在无需借助转速传感器的前提下从高精度时频表示中重构瞬时转速曲线；提取监测数据的瞬时特征，建立监测数据的低冗余稀疏表示和数据驱动的深度学习故障诊断模型，运用多目标优化方法提升模型的泛化性能。所建立故障诊断模型的特点是：可以在实验室内的定转速工况下完成训练，然后在实际变转速工况下实现故障部件的定位与诊断。本研究将为能源与动力装备的可靠运行提供技术支持，推动我国能源动力装备水平的进一步提升。

内容获取失败，请点击重试