New Multi-class Support Vector Algorithm and Its Application in Fault Diagnosis

Hierarchical support vector machines (H-SVMs) are faster in training and classifying than other usual multi-class SVMs, and therefore they are suitable for on-line fault diagnosis. A new multi-class fault diagnosis algorithm was proposed based on H-SVM. Before SVM training, the training data were first clustered according to their class center Euclid distances in some feature space. The patterns which have close distances were divided into the same sub-classes for training, and this made the SVMs have better generalization performance and reasonable hierarchical construction. Instead of common C-SVM, ν-SVM was selected as binary classifier, in which the meaning of parameter ν was more obvious and could be determined more easily. A simulation diagnosis experiment for the gas path components of a turbojet engine is conducted to demonstrate the effect of the algorithm. The simulation results show that the designed H-SVMs can fast diagnose 5 classes of single fault and 8 classes of combination fault for the engine. The fault classifiers have good accuracy and good generalization performance. As an application example, 6 kinds of real fault samples for JT9D engine were also classified correctly using the algorithm.