Variable Predictive Model Based Class Discriminate And Its Application In Fault Diagnosis Of Roller Bearing

Roller bearing is one of the common parts used in rotating machinery, but it is also one of the easily damaged components due to poor working conditions. Many faults of rotating machinery are related with roller bearing, and result in that rotating machinery is difficult to run properly. The failure of equipment could cause the problems of entire production line, and leads to economic losses and even casualties. Therefore, it has very important significance to monitor and diagnose the state of roller bearing.Essentially, the fault diagnosis of roller bearing is a pattern recognition process. A new method for pattern recognition-variable predictive model based class discriminate(VPMCD) is proposed by Raghuraj and Lakshminarayanan. The method establishes variable prediction models according to the inner relationship between the samples' feature parameters, and then to recognize and classify the testing samples through variable prediction models which had been established. Generally, in the fault diagnosis of roller bearing, there is a certain internal variable relationship between feature parameters; and the relationship is also different in different categories or systems. Therefore, the pattern recognition method-VPMCD can be applied in the fault diagnosis of roller bearing.In this paper, the fault diagnosis method of roller bearing based on VPMCD is proposed. The main research contents of this paper are as follows:1. The fault mechanism of roller bearings and its faults' vibration characteristics are discussed. And the fault diagnosis methods of roller bearing are briefly introduced from two aspects: signal processing and pattern recognition.2. The basic theory of VPMCD method is elaborated, and VPMCD is compared with ANN(Artificial neural networks) and SVM(Support vector machine) through UCI standard data and experiment data to prove its advantages. And the effectiveness of VPMCD is verified.3. VPMCD classification method is built on homoscedastic regression model, when the regression model is heteroscedastic, it will lower the prediction accuracy. Aiming at this defect, the parameter estimation approach, ordinary least square method, has been replaced by weighted least square method, which can get more precise model parameters, thus raise the accuracy of pattern recognition. Local characteristic-scale decompositio n(LCD) is applied into feature extraction, and blind source separation based on LCD is studied. The analysis results from the roller bearing vibration signals in different states demonstrate the effectiveness of the method of WVPMCD and LCD.4. Aiming at the defect that least square estimation will lower the robustness of model when predictive variables exist multiple correlations, the method of VPMCD based on partial least square is proposed to eliminate the multiple correlations and improve the prediction accuracy of the model. Adaptive and sparsest time-frequenc y analysis(ASTFA) is a new analysis method, and it translates signal decomposition into optimization problem, and it is very suitable for fault vibration signal processing of roller bearing. The two methods are applied into fault diagnosis of roller bearing, and the analysis results show the effectiveness of this method.5. Aiming at the problem of “dimensionality disaster” when dealing wit h multi-dimensional data using VPMCD, LDA(Linear discrimination analysis) is combined with VPMCD to be applied in the deterioration state identification of roller bearing, and is compared with PCA(Principle components analysis). High-dimensional feature vector is compressed to obtain a low-dimensional feature vector, which retains the nature of information and conforms to the modeling thought of VPMCD, so it is advantageous to the fault diagnosis online. Experimental results show that this method can be effectively applied to the deteriorated state identification of the roller bearing.