Abstract: A fault diagnosis method of hydroelectric generating sets based on multi-dimensional features and multiple classifiers is developed. Multi-dimensional features are constructed by extracting time domain characteristics, frequency domain characteristics, and sample entropy of ensemble empirical mode decomposition from the vibration signals of the generating units in different working conditions, and reduced by the genetic algorithm, so that this new method can achieve multidimensional information complementarity in the vibration features. With the multi-dimensional features as classifier inputs, faults are diagnosed using the support vector machine classifier, back propagation neural network classifier, and naive Bayes classifier. The preliminary diagnosis results of the three classifiers are fused to draw the final diagnosis conclusion, thus improving the accuracy of fault diagnosis of the generating sets. To verify the method, rotor unbalance, rotor misalignment and rotor rubbing are simulated experimentally on a rotor test bench, and the method is used to diagnose these faults. The results show that the diagnosis accuracy of multi-dimensional features and multiple classifiers is much higher than that of the single dimension feature and single classifier.

Key words: multi-dimensional features, multiple classifiers, sample entropy, hydroelectric generating set, fault diagnosis