To address the current situation that the fault feature dimensionality is too high， which leads to poor performance of fault classification and identification techniques， a fault dataset dimensionality reduction algo?rithm based on Median feature line multi-graph embedding （MFLME） is proposed. The algorithm improves the projection metric from sample points to feature space into the median metric in order to weaken the extrapolation error of the algorithm. Next， by defining the near-neighbor feature line graph and the far-neighbor feature line graph， the confusion of different kinds of samples is reduced and the category spacing is expanded， which re?duces the difficulty of the subsequent fault classification decision. Two different rotor failure simulation experi?ments are used to validate the algorithm performance. The results show that the algorithm can reduce the diffi?culty of fault classification and improve the identification accuracy.