Compound fault signals characterized by mutual coupling of symptoms, non-linear, and non-stationary, etc., it difficult to accurate and effective diagnosis. In this paper, a novel method employs transient impulse extraction and kernel principal component analysis (KPCA) to diagnose compound faults. The transient impulse extraction in the presented paper as a fault signal extraction method does not need prior knowledge of bearing vibration signal can effectively extract transient fault signal. The extracted signal employs frequency symptom parameters to express fault features. Kernel principal component analysis is a improve algorithm based on standard PCA. It can deal with the non-linear data. Through analyzing the distribution of principal component score can separate the single fault. The proposed method has merits as: (1) it differs from signal resonance demodulation method without the need to know the fault characteristic frequency; (2) is based on the transient impulse extraction of a single fault signal from compound signal can effectively separate and extract single fault features from compound fault signals. The effective and advanced performance of the proposed method is validated by the data of rolling bearing on the rotating machine experiment bench.