Rotating Machinery Failure Diagnosis Based On Support Vector Machine

Fault diagnosis is essentially a pattern recognition. The development of methods for improving the accuracy and effectiveness of the pattern recognition is always an interesting topic. It is particularly true for the fault diagnosis of rotating machines such as generators, turbines, pumps and compressors, as they are widely used in industry and are crucial basis for normal production in a plant. A number of condition monitoring and fault diagnosing systems of rotating machinery have been developed and applied in practice. The improvement and enhancement of these systems are however expected by both users and developers. Much work remains to be done.Statistical learning theory and support vector machine are interesting and promising technique applicable to improving the fault diagnosis of rotating machinery. In this dissertation, an investigation on the theoretical basis of support vector machine and its application in the fault diagnosis is carried out.The main works and contributions of this dissertation are summarized as follows:Statistical learning theory and support vector machine are briefly introduced. Analysis of the most suitable structures of support vector machine for application to the fault diagnosis is performed. Experiments are carried out with a test rig to examine the effectiveness of two typical kinds of artificial neural networks. Furthermore comparisons between the fault diagnosis based upon neural networks and that based upon support vector machine are taken. The results show that the fault diagnosis based upon support vector machine can be well trained with much less amount of samples than that based upon neural networks needs. Additionally, the optimum structure and parameters of the support vector machine can easily be determined by the learning process, however the neural networks can not.An information gain of signature signals is introduced to assess the contribution of the signature signals to diagnosing faults in rotating machines. For example, the information gain of 1x forward and backward whirls is high for detecting rotor imbalance and the gains of 1x, 2x and 3x forward and backward whirls are high for detecting stator/rotor rub. These results obviously correspond well to the existed practical experiences. Therefore the information gain is a suitable criterion for choosing signature signals.SV-Whirl graph is established to recognize rotor imbalance and stator/rotor rub. The experiment gives a satisfactory validation, that is, the imbalance and rub are clearly distinguished in the graph. Moreover, the severity of the faults can be visualized with colors and their brightness.