The invention relates to a failure prediction method of a roller bearing based on a partial least squares extreme learning machine. The method herein includes: analyzing feature indexes, such as timedomain, frequency domain and time-frequency domain, providing a feature extraction method based on the combination of half-normal distribution and empirical wavelet denoising to perform failure diagnosis on a roller bearing so as to obtain better denoising effect owing to proximity to original signals; for multi-feature parameters, comprehensively evaluating failure attenuation features of the roller bearing, and providing a method with the combination of residual-modified ISOMAP (isometric feature mapping) nonlinear feature dimension reduction and fuzzy C-means, so that change tendency and sorting precision are improved for the roller bearing in different attenuation stages; based on the extreme learning machine theory, providing a data prediction model based on a partial least squares extreme learning machine, optimizing parameters in the ELM (extreme learning machine), selecting node quantity of an optimal hidden layer and weight value of a connection layer, and selecting a Softmaxactivation function. Therefore, prediction precision is high, calculating time is short, and post-clustering feature value detection is effective. The failure stage of the roller bearing can be precisely predicted via the above steps.