Support Vector Data Description Of The Application In Fault Diagnosis

Support vector data description (SVDD) differs from the conventional pattern classification method. In conventional pattern classification problem, data from two classes (or multiple classes) is needed, the decision boundary is supported by examples from both sides. Most conventional classifiers assume more or less equally balanced data classes and do not work well when one class is severely undersampled or even completely absent. In SVDD only one class of data is needed to train a classifier and outlier can be distinguished from target objects. For some important equipment, fault is not allowed or very low fault rate is desired. Appling SVDD to mechanical fault diagnosis and condition monitoring. It is expected to solve the problem that conventional classification method meets when fault objects are very rare or absent.1. The basic theory and algorithm of SVDD are introduced. In order to make the classifer more flexible, kernel function is introduced to replace the inner product.2. Because the empirical mode decomposition has advantage in processing non-stationary signal, using this method in data preprocessing to extract the energy varies in different frequencies bands, and using the energy features to train the SVDD. The test result show that this method can preserve the features of original signals and get a good result.3. When few fault samples are available in fault diagnosis, it is possible to train both a traditional classifier(two-class classifier)and a SVDD. But SVDD just using normal samples, and the traditional classifier is expected to perform very poorly when just a few fault samples are available or extremely undersampled. An improved SVDD is researched in this paper: the SVDD with fault samples. Applying this method to the rolling bearing fault diagnosis, the test result shows that this method is superior to original SVDD method and would identify rolling bearing fault patterns more effectively.4. Three main principles of one-class classification methods is introduced: density estimation, direct boundary estimation and reconstruction. Their performance in dealing with different dataset is evaluated and compared. The density estimation gives the most complete description of the data, but might require too much data. For lower sample sizes a method which directly estimates the boundary might be preferred, especially the SVDD. Finally, a distance or a reconstruction error might be defined based on a model of the data. This model gives the ability to include extra prior knowledge of the problem.