Railway axles are critical elements as train loads and travel speeds have increased over time, requiring more efficient non-destructive inspection methods. Here a novel online procedure based on empirical mode decomposition (EEMD) and optimized support vector machine (SVM) using axle-box vibration signals for multi-state estimation of axles is presented. The axle-box vibration signals are adaptively decomposed into a number of intrinsic mode functions (IMFs) by EEMD. Two types of features, the relative EEMD time-frequency entropy (EtfE) and singular values of the matrix, whose rows are IMFs, are extracted and used to generate the feature vector, which would be input to the SVM to estimate the work state of axle. The proposed approach was applied on a wheel-axle system experiment with 7 kinds of defective axle. All axle-box vibration signals contain sixteen groups of axle working condition, all of which were calculated to evaluate the effectiveness of the proposed method. The results show that such feature combination and extraction based on EEMD entropy and SVM can identify the normal and 7 kinds of defective axle with high classification rates. The proposed approach outperforms in the practical wheel-axle system with strong background noise, and more importantly it can enable to perform online.