Fault feature extraction is critical in the fault diagnosis of avionics equipment. At present, Self-Optimization Kernel-Kernel Entropy Component Analysis (SOK-KECA) is an effective approach. But in consideration of the high computational complexity of SOK-KECA with large feature dimension, a two-stage feature extraction strategy is further proposed based on one-dimensional feature ambiguity. The improved minimum redundancy maximum relevance criterion is firstly used to execute feature rough selection, and then SOK-KECA algorithm is used to execute accurate feature extraction. Case study verifies that the proposed method can extract more discriminative features and have the ability to suppress noise in normal circumstances.