Considering the state parameters significant nonlinearity and the vulnerability to noise pollution in the aero-engine gas path faults，a method based on denoising autoencoder (DAE) and integrated with a neural networks of firefly algorithm (FA) and radial basis function (RBF) is proposed to diagnose the gas path faults and improve the diagnostic accuracy. The DAE is adopted through greedy algorithms to identify deeper robust features that helps diagnose the faults. To further improve the diagnostic accuracy of the algorithm, inertia weight and improved FA of selfadaptive light intensity factor are introduced to obtain the firefly radial basis function (FRBF) network after optimizing the RBF network. Then the robust features extracted from the DAE are imported into the FRBF for faults diagnosis. Based on practices, the extracting method is compared with the algorithms which are original DAE, independent FRBF, SVM and RBF. According to the results, the proposed method presents highest diagnostic accuracy of 98.1%, stable performance in the algorithms and more satisfying robustness.