Research On Fault Feature Extraction Methods Of Gear Transmission System Based On Higher Order Statistics

This thesis took the fault vibration signals of gear transmission systems as the study objects. On the basis of analyzing the vibration mechanism and the failure modes, vibration signals were divided into various components from the perspective of Gaussian characteristic and non-Gaussian characteristic. Extracted-non-Gaussian intensity features and bispectral entropy features from bispectral amplitudes of vibration signals. Finally discussed the issues of feature optimization and practicality of new higher order statistics characteristics. The main achievements and results of this thesis are as follows:(1) By studying the failure modes, the vibration mechanism and the fault performance of gear transmission systems, vibration signals could be devided into several components such as the non-Gaussian deterministic components with the relevent of fault parts, the non-Gaussian deterministic components with the irrelevant of fault parts, the Gaussian random components, the symmetrical non-Gaussian random components and the asymmetric non-Gaussian random components. When gear or bearing is fault, the deterministic components and the random components of vibration signals would significantly change, and the non-Gaussian characteristic of the signals also changed. Especially the sideband component which was enhanced due to the fault, would cause the enhancement of non-Gaussian characteristic and quadratic nonlinear characteristic. The bispectral analysis results of the signals were very sensitive to this phenomenon. In addition, bispectrum can suppress the Gaussian random components and the symmetric non-Gaussian random components of vibration signals, meanwhile retain the information of asymmetric non-Gaussian random components. This was conducive to reducing the interference of noise components and non-faul random vibration components.(2) When a gear transmission system goes out of order, the distribution intensity and morphology of non-Gaussian components of vibration signals will change in bifrequency domain along with the fault. Based on bispectral amplitude information, the non-Gaussian intensity feature and the bispectral entropy feature can be extracted. Non-Gaussian intensity was the quantitative description of intensity change for the non-Gaussian components of vibration signals, while bispectral entropy was the quantitative description of distribution morphology change in bifrquency domain for the non-Gaussian components of vibration signals. According to the different bifrequency definition domains,6types of features were extracted. The features were NGIPD and HB-PD which were based on the principal domain, NGIpart and HB-part which were based on a arbitrary two-dimensional interval called’part’, NGIregion(k) and which were based on multi-partition. These new higher order statistics characteristics not only retained the advantages of bispectral analysis, but also made up for the lack of conventional higher order characteristics such as bispectral slices. Different non-Gaussian intensity and bispectral entropy features had their own characteristics. NGIPD and HB-PD respectively contained whole intensity information and whole distribution morphology information of bispectral amplitudes, excluded bispectral symmetry redundant information at the same time. If the two-dimensional interval’part’contained much abundant fault information, the abilities of characterizing failures of NGIpart and HN-part were better, but it required human observation and a appropriate two-dimensional interval setting in bifrequency domain. As a uneven distribution of vibration signals’fault information in bifrequency domain, the fault features extraction of NGIregion(k) and HB-region(k) were better in some regions, and worse in other regions, which reflected that features based on bifrequency patition had strong abilities to focusing on bispectral contents and fault information.(3) In order to remedying the shortcomings of non-Gaussian intensity and bispectral entropy features, made two types of characteristics optimization which were feature compression and signal filtering.. The bispectral partition features had high dimensional feature spaces, and caused trouble in the actual application. Made feature compression by principal component analysis and kernel principal component analysis from the linear or the non-linear perspective, and it can concentrate most of the fault information into low dimensional principal component space or low dimensional kernel principal component space. Especially after the features prioritization pretreatment according to the Fisher criterion, the new principal component features and the new kernel principal component features had higher fault distinguishing abilities. The gear meshing frequency and its harmonic components usually had high energy and be vulnerable to non-fault factors. So utilized Gabor filter and signal reconstruction to accurately exclude the meshing frequency and its harmonic components of vibration signals in time-frequency domain. The results showed that the non-Gaussian intensity features of reconstructed signals were more sensitive to gear faults. This was conducive to the follow-up fault diagnosis. In order to research the usage methods and the effects of the new features applied in condition monitoring and fault alarm for gear transmission systems, took the HB-PD of original signals and the NGIPD of reconstructed signals passing Gabor filter as examples, made trend analysis of these fault features and set fault threshold values according to the’3σ criterion’. Both of the alarm time and the diagnostic accuracy obtained ideal effects. It provided new ideas and new methods for the practical application of engineering.