For compound fault features easy submerged by noise signal. Traditional time-frequency analysis and manifold learning method can't complete failure for efficient mining of potential information and further realize the fault feature extraction. Based on manifold learning is proposed on the basis of a manifold subband thought and its application to the study of compound rotor faults. Then draw a manifold subband feature mapping method based on compound rotor fault. First of all,to the phase space reconstruction fault original signal sequence. combination with wavelet packet strong inhibitory to noise and the signal the characteristics of high resolution. Reconstructing signal is decomposed into different frequency bands. The same frequency band of the same fault and many conditions are intergrated into the band matrix and estimate the intrinsic dimension. At last, by Laplace feature mapping algorithm based on the intrinsic dimension Subband dimension reduction obtained low dimensional feature vector and extract the information entropy, then further realized the fault feature extraction. Experiments show that: Compared with the classical local linear embedding and Laplace feature map algorithm, etc. Not only for single fault but also for compound fault manifold subband feature mapping algorithm more completely and effectively digged and extracted the characteristics.