Research On Feature Extraction For Fault Signals Of Rotating Components In Helicopters Based On Higher-order Statistics

Helicopters are becoming more and more important in modern wars, so their reliability and flying safety have attracted more attention in the military domain. Rotating components are very easy to get abnormal and often not redundant, so in-flight malfunctions can lead to catastrophic results. Thus it is much significant to try to detect such faults early to improve reliability and safety of helicopters. One of its key techniques is feature extraction of fault signals.Fault signals of rotating components in helicopters are often complex due to their complex structures and serious operating circumstances. Condition signals caused by impulses and modulations are often non-Gaussian and (or) non-stationary. Among these condition signals, the prominent fault signals show phase coupling, impulsive, cyclostationary and so on. Traditional methods are often based on Gaussian & stationary assumptions and less suitable to extract the effective features readily. So it is valuable to study how to utilize higher-order statistics to extract fault features of helicopter rotating components.Supported by the National Defense Advanced Research Projects, this dissertation takes the intermediate gearbox of one helicopter as a diagnostic object and addresses the problem of feature extraction of phase coupling, impulsive and cyclostationary signals characterizing faults of helicopter rotating components using higher-order spectra, higher-order time-frequency distribution and cyclic-statistics included in higher-order statistics theory. The detailed contents and innovative work can be summarized as follows.1. The main difficulty faced by feature extraction of helicopter rotating components is analyzed and the traditional methods are summarized. Then the advantages of methods based on higher-order statistics are given and several problems are pointed out.2. As to the problem of feature extraction of phase coupling signals characterizing faults of helicopter rotating components, the method based on higher-order spectrum (HOS) is deeply studied and improved. At first, one 2-D Hanning-Poisson combined lag window is presented, which can improve the performance of bispectrum estimation. Secondly, a novel method of feature extraction using trispectrum is discussed. Finally, 21/2-D spectrum is used for on-line application of trispectrum.Numerical and experiment analysis demonstrate that bispectrum and trispectrum candetect phase coupling features of fault signals under background noises. Trispectrum can also be used to analyze signals with symmetrical probability density compared with bispectrum.3. The method based on higher-order time-frequency distribution is deeply studied to extract features of impulsive signals characterizing faults of helicopter rotating components. One two-step filtering preprocessing method is proposed to enhance weak impulsive signals, and two kinds of image features of impulsive signals are defined using image-processing techniques, which can be used to extract and quantify impulsive features automatically.The above research shows that two-step filtering preprocessing can suppress both narrow-band interfere signals and broadband stochastic noises, so it will increase the SNR of impulsive signals. Based on it the sliced Wigner trispectrum is feasible to extract impulsive features of faults automatically and its effectiveness is validated by early detection of one gear-pitting fault.4. In order to extract features of cyclostationary signals characterizing faults of helicopter rotating components, the methods of extracting fault features using cyclic statistics are deeply studied.(1) One novel method of extracting 1-D cyclostationary features of faults is presented based on spectrum line regeneration (SLR).(2) The way of extracting 2-D cyclostationary features of faults based on spectrum correlation density (SCD) function is discussed.(3) The idea of extracting 3-D cyclostationary features of faults using cyclic bispectrum is studied. Also its estimation algorithm based on 2-D Chirp-Z transformation is presented.Numerical and experiment analysis demonstrate that cyclic statistics can extract cyclostationary features close related to faults under background noises.5. Combined with the cyclostationarity of signals of helicopter rotating components, one new method of feature extraction based on cyclostationary time series model is proposed. One linear almost periodically time-varying AR (LPTV-AR) model is presented and the criterion of ordering the model is defined. Then the algorithms to identify model parameters are put forward in both time domain and frequency domain.Numerical and experiment analysis demonstrate that the method is insensitive to additive stationary noise and can be used to detect and predict "novel" abnormal conditions of helicopter rotating components.