Application of support vector machine to fault diagnosis of rotation machinery

An investigation into the theoretical basis of support vector machine (SVM)and its application to detect unbalance and rotor/stator rub in rotating machinery is carried out on a test rig. An experimental comparison of SVMs respectively based on two kernel functions, polynomial and radial basis functions, is made, and different signature quantities of vibration signals are inputted into SVM as source information. The results show that the optimum accuracy of fault diagnosis by both SVMs is almost identical and the performance of SVMs lessly depend on the structures (kernel functions), which makes SVMs easier to be applied in practice. However, the selection of signature signals inputted into SVMs as training data influences the accuracy of fault diagnosis markedly. For detecting unbalance and rotor/stator rub, 1x, 2x and 3x forward and backward whirls are the optimum signature signals. Additionally, the forward and backward whirls can be used to constitute SV-Whirl Graph to recognize rotor unbalance and stator/rotor rub clearly and visually.