Research And Application On Fault Diagnosis Of Rotating Machine Based On Variable Predictive Model Based Class Discriminate

It is essential to precede pattern recognition for fault diagnosis of rotatingmachinery, the neural network and support vector machine (SVM) are the widely usedand mature pattern recognition methods, but there are some inherent defects in thetwo methods. Especially, it is noteworthy that these pattern recognition methodsignore the inherent relationship between the extracted features. However, certaininternal relation exists among the features mostly in extracted characteristics ofmechanical fault vibration signal, and the internal relation is different obviouslybetween different systems and categories (the same system in different workstates).Based on this, a new method for pattern recognition based on Variablepredictive model-based class discriminate (VPMCD) is proposed by Raghuraj andLakshminarayanan, and it has been applied in the pattern recognition field of biology.The essence of the method is to establish mathematical model through the mutualinterrelationship between the characteristic values, and to get different mathematicalmodels for different categories, thus these mathematical models can be used to predictthe characteristic values of test samples, and the predicted results could be employedas a basis for necessary classification and for further pattern recognition.Theestablished nonlinear model can solve the pattern recognition problems of nonlinearsystem that are described by the multivariable, therefore the VPMCD method isintroduced into fault diagnosis of rotating machinery in this paper, which will be useto classify working condition and fault types of rotating machinery, and morecomprehensive research proceeded for the VPMCD itself, then the improved VPMCDmethods are put forward. The experimental analysis results show that VPMCD andimproved VPMCD method can be effectively applied to fault diagnosis of rotatingmachinery.The main research contents of this paper are as follows:1. The widely used pattern recognition methods are introduced, and the basicclassification principle and the main advantages and disadvantages are discussed, thenbased on that, the Variable predictive mode based class discriminate is introduced.The feasibility of the application for this method is proved by discussing the basicprinciple and application of VPMCD.2. VPMCD method is applied to fault diagnosis of rotating machinery, and combining with Local characteristic-scale decomposition (LCD) de-noising method,fuzzy entropy feature extraction method. By analyzing the experimental data,compared with the pattern recognition methods such as neural network and supportvector machine (SVM), the iterative learning of neural network and the optimizationprocess of SVM are avoided, and the operand is greatly reduced, then the time ofclassification is shortened in VPMCD. Moreover, the establishment of the predictionmodel is the process of parameter estimation essentially in VPMCD, which will avoidthe choice of structure and type in neural network, and the selection of kernel functionand parameters in SVM will be avoid. Therefore, VPMCD pattern recognition methodis less influenced by subjective factors, and the classification results are moreobjective and more accurate.3. Targeting that the extraction features contains irrelevance, smaller correlationand redundancy features, which will be selected by distance evaluation technique(DET) method based on Wrapper model, and the irrelevance, smaller correlation andredundancy features are remove, therefore, the most sensitive features to the class arechoose. In addition, least-squares regression is used to fit parameters in VPMCD,however the least-squares method itself has some inherent defects, so the establishedprediction model cannot meet the requirements of classification eventually, Aiming atthe defects of the least squares regression, the robust regression is used to estimateparameters instead of least-squares regression in this paper, which can avoid thedefect of the least-squares, thus more accurate prediction model can be got.4. Stepwise regression is applied to VPMCD for pattern recognition, introducingvariable and calculating the significant level by stepwise regression, to establish theprediction model that only contains significant characteristic values, achieving thefunction of embedded feature selection and modeling simultaneously; Moreover,compared to neural network and support vector machine, the greatest advantage ofVPMCD is the timeliness when the characteristic value is small, meanwhile, thebiggest drawback of VPMCD is low efficiency when the dimension of characteristicsis high, which will be bad for online diagnosis, VPMCD based on stepwise regressioncan effectively solve the impact to VPMCD when feature dimension is more.5. Laplacian Eigenmaps (LE) and Kriging function are applied to VPMCD forpattern recognition, the extracted high-dimensional features are compressed by LEmanifold learning algorithm to obtain low dimensional characteristics with intrinsicregularity, which conform to the modeling principle of VPMCD exactly and retain thenature of information, so it is advantageous to the fault diagnosis. In addition, the Kriging function is applied to VPMCD, when the relation of data in feature set isrelatively complicated, four regression model in original VPMCD is difficult toaccurately establish prediction model, but regression model is used mainly and thecorrelation model is adopted complementally in VPMCD method based on Krigingfunction, thus more accurately prediction model is established. Therefore, LEmanifold learning, Kriging function and VPMCD are combined together, which caneffectively extract characteristics, establish models and precede pattern recognition.6. The incremental and semi-supervised theory are used in VPMCD for patternrecognition, the pseudo labels are provided to participate in VPMCD training throughthree stages of the study in the incremental semi-supervised VPMCD patternrecognition method, which can effectively solve the defect of the diagnosis on thecondition of small samples; VPMCD could make full use of the useful information inthe unknown samples through semi-supervised learning, thus make the establishedprediction model more real and improve the recognition accuracy effectively. Inaddition, the established prediction model that uses small samples belongs topretreatment stage before diagnosis, and pseudo scalar samples are introduced intoincremental semi-supervised VPMCD to update the model, which greatly shortens thetime of classification and thus provides online diagnosis in possible. The verifiedresults show the effectiveness of the method through theoretical analysis, simulationdata and experimental data.