When the traditional extreme learning machine is used to predict the rolling bearing fault, there is a problem that the original signal pattern is aliased, and the artificial parameter selection causes the prediction accuracy to be low, and the fault prediction method of the normal distribution-empirical wavelet transformation combined with partial least squares based on the extreme learning machine method is proposed.Firstly,the normal distribution-empirical wavelet transformation signal de-noising method is proposed. The normal distribution is used to determine the interval number to divide the frequency band boundary. A band-pass filter is constructed and de-noised on each partition interval. Secondly, the fault prediction method of PLS-ELM is proposed, the principal component number and load weight of the partial least squares method are applied to improve the number of hidden layer nodes and the network weight of the extreme learning machine respectively. The activation function selects Soft max to improve the fitting accuracy of the data.Finally, the kurtosis of dimensionless index is used to reflect the fault degree and realize the fault trend prediction.The experimental results show that the method overcomes the problem of modal overlap and realize the prediction of performance deterioration trend of rolling bearing.