Study On The Feature Extraction And Diagnosis For Surface Damage Of Rolling Element Bearings

The rolling element bearings are the commonly used in various rotary machinery, and their health conditions directly affect the production quality and safety of the machinery equipment and even the related complete fabrication line. Fault diagnosis and condition monitoring for rolling element bearings still remain a very important research and develop field. The vibration signal-based analysis is the most commonly used method in the bearing fault diagnosis, which will also be employed in this work.The noise reduction of vibration signal is very important to effectively reveal fault information; in this work, the de-noising method based on a comprehensive threshold kurtosis will be studied. In the process of bearing fault diagnosis and monitoring, the fault feature extraction and pattern recognition are two critical steps. The image-based feature extraction method will be investigated according to the characteristics of collected bearing signals. Several techniques are studied in pattern recognition and optimization for bearing fault detection, which include the Morlet wavelet transform, the least square support vector machine (LSSVM) and fuzzy c-means (FCM). Main research work and the related contributions are summarized as follows:(1) In signal de-noising, the commonly used wavelet de-noising methods are analyzed, and a comprehensive threshold method based on the kurtosis of vibration signal is proposed. The threshold is determined based on the wavelet coefficients of signal at different scales. A new function with adjustable threshold parameter is proposed by integrating the merits of soft and hard threshold functions. A comparision analysis is taken to examine the performance of the proposed technique with other commonly used wavelet de-noising methods.(2) In fault feature extraction based on image recognition, based on the mthods of Hilbert envelope and bi-spectrum, the bi-spectrum of the signal envelope is introduced. It is proposed to use the moments statistics of bi-spectrum to recognize the health condition of the bearings. Next the moments statistics are further processed by the principal component analysis to generate 4 principal components that are used as the input vectors for fault pattern recognition. A comparative analysis shows that the moments statistics technique outperforms other related feature extraction methods in the accuracy of fault pattern recognition.(3) The impacts of Morlet wavelet parameters on time-frequency analysis are analyzed. According to the fact that bearing defect will modulate the vibration signal properties, the feature extraction method based on the optimal Morlet wavelet is proposed. To achieve the optimal match for wavelet analysis with impulse feature of the signal, the objective function is selected as the product of the maximum wavelet coefficients with maximum value of the kurtosis of wavelet coefficients. The simulated annealing algorithm is adopted to optimize the two shape parameters of Morlet wavelet:bandwidth and center frequency. The effectiveness of the proposed method is verified by both simulation and experimental tests.(4) In fault pattern recognition, with regard to the difficulty to obtain a large number of typical fault samples in the engineering practice, the LSSVM method is introduced for intelligent fault diagnosis for rolling element bearings. The parameter optimization of the regularizing variableλand the kernel widthσis performed by using the simulated annealing algorithm, whereas the sensitive subset of features is determined simultaneously. To verify the effectiveness of this method, some bearing tests are taken under four bearing conditions, five different shaft speeds and two load levels. Fifty two features in total are extracted from the bearing signals. The recognition accuracy of this method is proved.(5) The weighted FCM algorithm based on class separability is proposed. The weights are calculated according to the class separability, and then weights are assigned to the corresponding features to reflect the sensitivity of the features to real bearing fault patterns. The fault diagnosis is made through bearing vibration signals,under four kinds of load levels and several defect types.Finally some conclusion comments from this research work as well as the possible future research topics are summaried in the end of this dissertation.