A fault diagnosis method for turbofan engine bearings based on cyclical extraction-reconstruction and TMFF-CNN

To tackle the issues of inadequate extraction of weak fault features amid intense noise interference and the limited ability of single spectral features to characterize fault information in the fault diagnosis of turbofan engine rolling bearings, a fault identification method for aircraft engine based on cyclical extraction reconstruction and tri-modal frequency fusion-convolutional neural network (TMFF-CNN) is proposed. The method first obtains signal node components through wavelet packet decomposition, fuses kurtosis values and correlation coefficients to construct a comprehensive parameter, and cyclically extracts and reconstructs high signal-to-noise ratio (SNR) signals as signal preprocessing. Subsequently, spectrogram, Mel spectrogram, and Mel frequency cepstral coefficient features are extracted from the reconstructed signals, and multichannel input is constructed through TMFF. Lastly, a CNN is designed to achieve fault classification. Experimental verification using Case Western Reserve University bearing datasets shows that compared with methods without preprocessing, using single-index reconstruction, or adopting single spectral feature as input, the proposed method achieves a 100% recognition rate in noise-free environments and maintains a recognition rate of 99.034% at a 10 dB low SNR. Additionally, fault simulation experiments on a turbofan engine intermediate bearing are conducted to validate the effectiveness of the proposed diagnostic approach in practical complex noise environments. The findings demonstrate that this method can effectively suppress noise interference and improve the ability of feature representation. Therefore, this method can serve as a reliable solution for identifying complicated signals from turbofan engine rolling bearings.