For rotation mechanical devices, vibration signals in startup and stopping are rich in fault information. However, under these conditions, the shaft speed changes rapidly, inflicting unstable features. Therefore, the TFR is blurred by obvious frequency modulating. Order tracking is proposed to provide stable signals in angular field for traditional diagnosis, but it requires high accuracy on resampling, meaning more computing time, thus unfits for online diagnosis. To utilize resampled signals, this article combined order tracking with machine learning, transferred the signals to extreme learning machine (ELM) for classifying, remedied the inaccuracy of resampling with the adaptability of machine learning. Short-time Chirp Fourier Transform (STCFT) gained the TFR and then IF. Integration of IF provided the phase information for resampling. By VMD, resampled signals turned into modes. These modes were then learnt and classified by the ELM. Adaptation of VMD in this method can overcome the modulate mixing, as well as accelerate the analysis; the ELM can effectively classify the fault features in modes, and comparisons with other methods show that the method is promising on online diagnosis.