基于卷积自编码神经网络的航空发动机轴承故障诊断方法研究

航空发动机轴承早期故障多是由于裂纹、疲劳剥落和保持架损坏造成的,这类型的故障在发动机振动信号中均会产生瞬时的冲击;但是,在早期故障中,振动信号由于夹杂过多部件耦合激励,缺陷冲击信号很难辨识,早期故障诊断十分困难;采用了基于卷积自编码网络的航空发动机轴承早期冲击故障特征提取方法,通过分析信号中冲击成分的周期性,利用卷积自编码网络的平移不变学习特性,自动捕获信号中的周期成分,将信号分解为由卷积核重构的多个特征分量,实现信号特征分量的自学习,考虑到峭度指标对信号冲击成分描述的特点,使用峭度指标作为最优特征分量的选取指标,进而实现早期冲击故障特征的提取;最后利用仿真数据和轴承数据验证了该方法的有效性.

Research on aero-engine bearing fault diagnosis method based on convolutional auto-encoding neural network

Early failures of aero-engine bearings are mostly caused by cracks, fatigue spalling and cage damage. These types of failures will produce instantaneous shocks in engine vibration signals. However, in the early faults, the vibration signal is excited by the coupling of too many components, and the defect impact signal is difficult to identify, and the early fault diagnosis is very difficult. This paper proposes a feature extraction method for early impact faults of aero-engine bearings based on convolutional self-encoding networks. By analyzing the periodicity of the impact components in the signal, using the translation invariant learning characteristics of the convolutional autoencoding network, the periodic components in the signal are automatically captured, and the signal is decomposed into multiple characteristic components reconstructed by the convolution kernel to realize the signal characteristics The self-learning of the components takes into account the characteristics of the kurtosis index describing the impact components of the signal, and the kurtosis index is used as the selection index of the optimal feature components to realize the extraction of early impact fault features. Finally, simulation data and bearing data are used to verify the effectiveness of the method.