Multi-fault diagnosis for rolling bearing based on double parallel extreme learning machine & kurtosis spectral entropy

In this paper, double parallel feedforward extreme learning machine (DP-ELM) model and kurtosis spectral entropy (KSE) are proposed for the early weak multi-fault diagnosis for rolling bearing. Extreme learning machine (ELM) has been widely used as a fast classification model. But a key problem for ELM is to reduce the impact of random variables and it can't make full use of the direct and indirect information among different layers. To solve the problems, a connection is built between the input layer and the output layer. That is, the output layer receives information not only from the hidden layer but also directly from the input layer. And a set of the input layer weights and the hidden layer thresholds will be chosen analytically by founding the best accuracy. Meanwhile, feature extraction is the prerequisite for classification. In order to improve the accuracy of the multi-fault classification, this paper applies the kurtosis spectral entropy (KSE) algorithm to get the useful features from the vibration signals. Then the eigenvalues are input into the DP-ELM model for pattern recognition. In this paper, two sets of rolling bearing data at different speed are used to test. And each set of data includes six different states. Compared with ELM, back propagation neuron network (BP) and radial basis function (RBF), the experimental results show that, with as few as possible of the hidden nodes, the improved method used in this paper ensures a short learning time that it might learn hundreds of times faster than BP method, and obviously increases the accuracy at a degree. Finally, 10 times 10-fold-cross-validation are used to prove the effectiveness.