Research On Rotating Machine Fault Diagnosis And Prediction Method Based On Support Vector Machine

With the rapid development of modern science and technology, rotating machinery continues melting towards maximization, complex, high speed, continuous process and automation directions. Not only do these developments lead to higher productivity but also higher requirements are put forward for the safe operation of the rotating machinery, which once goes wrong, enormous economic losses, even catastrophic casualties and serious social impact will be caused. Diagnosing the fault of the equipment and predicting its running state according to historical data are two important measures to ensure its safe and reliable operation. Only in this way, the existing abnormal problems will be detected timely and effectively, so that the fault can be eliminated in the bud. Hence, in this thesis, the two above measures based on support vector machine theory were investigated, a model of fault diagnosis and state prediction was set up, and a further study went on the solution to some key problems. The main achievements consist of as following aspects:1. Put forward a sample balance method of SVM based on genetic algorithm.For fault diagnosis, fault samples are generally less than the normal samples, so the problem of imbalanced samples universal exists. Using SVM to analyse the imbalanced samples often results in the misdiagnosis of less sample class. Thus, a sample balance method based on SVM of genetic algorithm were put forward to solve the problem, which can be explained by using the crossover and mutation of genetic algorithm to generate progeny sample, breeding and extending the less sample class to obtain more samples, and finally reaching the balance of the two kinds of samples. In order to make the extended samples more pertinence and more conducive to form the correct classification hyperplane, selection method of the parent sample and evaluation method of the progeny sample were given.2. Analysed the causes of illusive component resulted from EMD, and proposed a method to identify the illusive component.As a feature extraction method, EMD can not only deal with the non-stationary and non-linear problem preferably but to often introduce illusive component by using it for signal processing, which affects the accuracy of the analysis and seriously restrains the development of EMD. For the sake of eliminating the effection of the illusive component and preferably playing EMD in the role of feature extraction, an effective method to identify illusive component was proposed, which used K-L divergence to evaluate the degree of relationship between the decomposition components of EMD and original signal. The smaller K-L divergence the greater degree of relationship as there was higher authenticity of the component, conversely higher falsity. The illusive component can be distinguished through the threshold setting, and the method of which was given.3. Raised SVM algorithm based on EMD feature extraction.Correlation degree of the characteristic parameter and the predicted point, to a great extent, determines the accuracy of the predicted value while using historical data to predict the running state. At present, there are mainly two selection methods of characteristic parameters:one is to take the measured data, which are the influencing factors associated with predicted value, as the characteristic parameter, such as the wind speed and the air pressure are regarded as the characteristic parameters on predicting the power of wind power, but some pre-measure for vibration, due to the influencing factors are often very complex and difficult to clear, a high-precision prediction model will not be able to be created by using this method; the other is that the characteristic parameters are obtained by calculating the time series, and the most representative and most commonly used one is the method based on phase space reconstruction, which makes use of chaos theory to calculate embedded dimension and time delay for reconstructing phase space and consequently obtaining the characteristics of the time series. However, the determination of embedded dimension and time delay were only considered from the view of dynamics, the characteristic parameters obtained by what is not necessarily appropriate for the prediction model. It is for the inappropriate selection, the appropriate characteristic parameters can not be obtained, resulting in the prediction accuracy greatly reduced. For the problem of feature extraction of prediction model, SVM algorithm based on EMD feature extraction (EMD-SVM) was raised, which took the component value of each time point decomposed from EMD as characteristic parameters and together with the time series value (target value) constituted the sample to establish the prediction model, and its high stability and accuracy were proved by experiments.4. Aim at the problem of large-scale training sample of SVM, a new length selection method of the sample based on information entropy was put forward.The problem of large-scale training samples has always been a bottleneck plagued the computing speed promotion of SVM. Excessive training samples will greatly increase the calculation cost, not necessarily lead to more accurate prediction results, and even result in more serious deviation. Therefore, the length of the training samples must be controlled within a proper range. Aiming at the above problems, a new length selection method of the sample based on information entropy was put forward. The main idea of the method is to choose the historical data, which most relevant to the prediction point, as the training samples to guarantee the completeness and non-redundancy of the data information. The stronger the relevance of historical data and prediction point is, the larger significance the data as the training samples for forecasting point will be; it goes the other way around with the increasing distance (time goes forward), under the premise of adding the above date to the training sample, the volatility increases and the stationarity reduces, thus prediction point need to be abridged. In this thesis, the time series were cut out through selecting different position from front to back in turn as a starting point in the historical data, of which information entropy were calculated. The starting point corresponding to the minimum information entropy was regarded as the0point coordinate of the new time axis. The negative half shaft data were abridged because of the remoteness and the low degree of correlation with the current state, and due to the high degree of correlation, the positive half shaft data were taken as the training samples, thus completeness of the information can be guaranteed, at the same time the consideration was avoided for the lower related degree training sample in the optimization process of modeling. From the view of computation time and prediction accuracy, the validity of the method was investigated by means of theoretical analysis and experiment.