| As an important support component in aero-engine, the operating condition of rolling element bearing has a significant impact on the safety and reliability of aero-engine. Traditionally, time-based maintenance is performed to ensure the normal operation of rolling element bearing in aero-engine. However, owing to the large dispersion of rolling element bearing life, the mode of time-based maintenance is likely to result in disrepair or excessive maintenance. Therefore, it is necessary to replace traditional time-based maintenance by condition-based maintenance. In order to carry out condition-based maintenance for rolling element bearing in aero-engine, perfect condition monitoring and fault diagnosis technology is needed. In this thesis, based on advanced signal processing methods and kernel methods, condition monitoring and fault diagnosis technology for rolling element bearing in aero-engine is studied from three aspects: the sensitivity of fault diagnosis based on the signal measured on the casing, fault detect and fusion fault diagnosis. The main contributions of the thesis are summarized as follows.(1) An intelligent fault diagnosis approach for rolling element bearing based on regularized kernel discriminant analysis(RKDA) is proposed. In this approach, first, the envelope spectrum features and the energy features are extracted, respectively, from the vibration signals of rolling element bearing by using wavelet packet transform. Then, the two kinds of features are used as the fault features of rolling element bearing. Second, the extracted rolling element bearing fault features are projected onto a low-dimensional subspace by RKDA, in which the rolling element bearing samples of different conditions will have the best discrimination. Finally, rolling element bearing condition recognition is carried out by the nearest neighbor method in the subspace. The experiments show that the fault diagnosis accuracy of the proposed approach is similar to that of the approach based on support vector machines and this approach can be used as a good intelligent fault diagnosis approach for rolling element bearing.(2) The sensitivity of fault diagnosis of aero-engine rolling element bearing based on the signal measured on the casing is analyzed. The impact response test and the fault simulation test of rolling element bearing are performed, respectively, by using two aero-engine rotor test rig with casing. In the impact response test, the response of the impact from the location of rolling element bearing is first measured both on the bearing housing and casing, and then the differences between them are compared and analyzed in detail. In the fault simulation test of rolling element bearing, the signal of rolling element bearing is first measured both on the bearing housing and casing, and then the time domain waveform, frequency spectrum and wavelet envelope spectrum between them are compared and analyzed in detail. The results show that the joint stiffness between rolling element bearing and casing has great effect on the impact feature in the signal measured on the casing. When the joint stiffness is low, the rolling element bearing fault impact of the signal measured on the casing will be attenuated significantly. By using the wavelet envelope analysis approach, however, the fault of rolling element bearing still can be diagnosed. This research provides an experimental basis for the actual fault diagnosis of aero-engine rolling element bearing based on the signal measured on the casing.(3) A fault detection approach for aero-engine rolling element bearing based on a small sphere and large margin approach is proposed. When fault detection of aero-engine rolling element bearing is performed via machine learning, there are three cases of training samples, namely, only normal samples, the number of normal samples and fault samples is highly unbalanced, and the number of normal samples and fault samples is balanced. Previous studies only focus on the first and the third cases, and the fault detection approaches for these two cases are also different. Aiming at this problem, a fault detection approach for aero-engine rolling element bearing based on a small sphere and large margin approach is proposed in this thesis. This approach provides a uniform solution to the fault detection under different cases of training samples.(4) Two fusion fault diagnosis approaches for aero-engine rolling element bearing based on multiple kernel support vector machines are proposed. When fault diagnosis of aero-engine rolling element bearing is performed based on the signal measured on the casing, the fault features of rolling element bearing in the signal measured on the casing are very weak. If only a single type of fault features or data from a single sensor are used for fault diagnosis, the high diagnosis accuracy is usually hard to be obtained. Aiming at this problem, in this thesis, a multi-feature fusion diagnosis approach and a multi-sensor data fusion diagnosis approach based on multiple kernel support vector machines are proposed, respectively. The experiments show that, comparing with the approaches that only use a single type of fault features or data from a single sensor, the two proposed fusion diagnosis approaches can improve the diagnosis accuracy significantly. In addition, since the two fusion diagnosis approaches avoid the problem that kernel matrix needs to be selected by user, the automation level of fault diagnosis is improved further.(5) Two fusion fault diagnosis approaches for aero-engine rolling element bearing based on regularized multiple kernel discriminant analysis are proposed. Considering that regularized kernel discriminant analysis can achieve similar classification performance with support vector machines, in this thesis, the application of regularized multiple kernel discriminant analysis in the fusion diagnosis of aero-engine rolling element bearing is further studied. First, in order to use multiple types of features to improve the diagnosis accuracy, a multi-feature fusion diagnosis approach based on regularized multiple kernel discriminant analysis is proposed. Second, in order to use multiple sensor data to improve the diagnosis accuracy, a multi-sensor data fusion diagnosis approach based on regularized multiple kernel discriminant analysis is proposed. The experiments show that, comparing with the approaches that only use a single type of fault features or data from a single sensor, the two proposed fusion diagnosis approaches can improve the diagnosis accuracy significantly. In addition, since the two fusion diagnosis approaches avoid the problem that kernel matrix needs to be selected by user, the automation level of fault diagnosis is improved further. |
