Abstract: In order to improve the fault recognition accuracy of rotating machinery, an ensemble RBF fault pattern recognition method combining neural network and ensemble learning method is proposed. First, the extracted rotor fault feature dataset is used to calculate the weights of each feature by ReliefF algorithm, which are arranged in descending order to filter out the series of features with larger weights to form a low-dimensional feature dataset. Secondly, the larger weights are used as the input of the non-relaxation roulette algorithm to perturb the series low-dimensional feature dataset and divided them into training sets and testing sets by disturbing corresponding low-dimensional data datasets. Then, the new training set is perturbed by the self-sampling method in the Bagging algorithm to train the corresponding number of RBF neural networks to further complete the construction of the discrepancy subclassifier. Finally, the test data identification results of each neural network are combined with the relative majority voting method to obtain the fault identification results. The experimental results show that for the fault identification of rotor system, the proposed method has higher identification accuracy and better generalization performance compared to nonensemble RBF neural network and ensemble BP neural network.

Key words: ReliefF lgorithm, ensemble learning;RBF neural network, feature disturbance, sample disturbance