Research On Feature Extraction And Condition Assessment Method Of Rolling Bearing Vibration Signal

Rolling bearing is the key component of many rotating machinery a nd is called the machine joint. Under extreme circumstances and various influence factors, rolling bearing component has the worst reliability of entire rotating machinery system and is ―Bucket Short Board‖ which directly influences the running reliability of entire machinery device. While roling bearing is running, its performance will degenerate from normal condition until final failure. So, if the performance degradation degree or fault location and performance degradation degree can be detected simultaneously during the performance degradation processing of rolling bearing, the traditional periodical or breakdown maintenance will be changed as condition-based maintenance. Once Predict and Prevent(PAP) mode is achieved, bearing lifetime can be used as long as possible, the maintenance cost can be reduced, more severe accidents and great loss will be avoided. This research direction focuses on device entire performance research and deeply expands the present fault diagnosis technology on the theory and met hod ways.In the paper, rolling bearing is the research object and the nonstationary vibration signal of rolling bearing is analyzed and processed. This research is targeted at a different conditions(normal condition, inner raceway, rolling element, outer raceway fault condition and different performance degradation degrees) intelligent assessment method of rolling bearing, on the basis of deeply researching feature extraction and reduction method of vibration signal, further researches the optimal problem of hyper-sphere-structured multi-class support vector machine and the key technology problem that how to build rolling bearing multi-condition assessment model. The main contents of the paper include:1. Research on time-frequency analysis method of vibration signal based on ensemble empirical mode decomposition(EEMD). Compared with other methods, Hilbert spectrum time-frequency analysis method based on EEMD is proved and shows some advantages such as higher time and frequency resolution and anti-mode aliasing. Aiming to the selection problem of two important parameters that adding noise amplitude value and overall average time in EEMD, the criterion problem is researched for adding white noise in EEMD method from the aspect of energy standard deviation. Rolling bearing vibration signal is decomposed by improved EEMD, intrinsic mode function(IMF) components can be obtained, but there are still pseudo components and insensitive components for rolling bearing fault, so an IMF extraction algorithm is studied by using kurtosis value combined with normalized correlation coefficient. Experimental research verifies the proposed method is effective and laies a good foundation for follow-on further feature extraction.2. Research on multi-domain feature extraction and reduction method of vibration signal. For subtly describing running condition of rolling bearing and reflecting global feature and local feature of rolling bearing vibration signal, multi-domain feature extraction scheme is researched which contains tim e domain, frequency domain, time-frequency domain. Therein, for time-frequency domain feature extraction method, several methods are proposed based on improved EEMD sensitive IMF respectively combined with time domain indexes, frequency indexes, autoregression(AR) model and singular value decomposition(SVD) method. Then, feature vector and each condition feature vector matrix of rolling bearing single sample are constructed and each condition feature library of rolling bearing can be obtained. Aiming to the relativity and redundancy problems among high dimensional features, manifold learning algorithm is researched. Combining with support vector machine(SVM), through experimental comparison analysis, the most effective feature reduction method for rolling bearing is determined.3. Research on intelligent classification method and fault intelligent diagnosis method. Although hyper-sphere-structured multi-class SVM has a series of advantages, its classification precision is not higher than ordinary SVM. Aiming to this problem, the classification rules are studied and the improved scheme is proposed, and new decision criteria are present for the key region. Else, for the problem that kernel parameter selection range of h yper-sphere-structured multiclass SVM is decided by experience, the distance calculation formula between hypersphere centres is deduced and regarded as separation index to decide the optimal selection range of kernel parameter, and achieves the training time loss. An intelligent fault diagnosis method is deeply studied for different running conditions of rolling bearing, intelligent diagnosis model of h yper-sphere-structured multiclass SVM is constructed, and by large number of experimental analysis, the validity of the proposed method is verified.4. Research on condition assessment method of rolling bearing. For only the subordinate relationship of bearing fault conditions can be judged by rolling bearing fault intelligent diagnosis method, damage extent and fault variation can ’t be quantitatively described and its performance condition problem can ’t be quantitatively assessed. By analyzing from SVM classification principle, rolling bearing structure and vibration propagation mechanism of sensor installation position, the assessment index that compensated relative distance is proposed based on SVM. By analyzing from improved hyper-sphere-structured multi-class SVM principle, the direction of feature vector and the position relations of each condition hypersphere, the assessment index that angle cosine distance compensation generalized minimum distance is proposed, and the intelligent assessment model is founded. By experimental research of incomplete vibration data and lifetime complete data of rolling bearing, the performance of each assessment model is compared and analyzed.