RTFNN: A refined time-frequency neural network for interpretable intelligent diagnosis of aero-engine

Convolutional neural networks (CNN) have received extensive attention for their powerful fault feature mining ability, which is widely applied in intelligent fault diagnosis and remaining useful life prediction of machinery. However, the powerful nonlinear mapping ability of CNN to map fault feature information to high-dimensional feature space leads to uninterpretable phenomena in their intelligent diagnosis results. Therefore, it restricts the application of CNN in the intelligent diagnosis of aero-engines with high reliability requirements. In order to address the aforementioned issues, an interpretable intelligent diagnosis method for aero-engines with refined time-frequency neural networks (RTFNN) is proposed. Since the aero-engine usually under nonstationary operating condition and strong noise interference, the traditional time-frequency analysis method cannot perform the optimal time-frequency representation. Therefore, a higher-order phase operator is constructed to transform the convolution kernel in the CNN into a refined time-frequency convolution kernel, and its parameters are iteratively optimized during the training process of the RTFNN model. Thereby, the strong noise interference component in the aero-engine vibration signal is effectively suppressed. Besides, the interpretability of the proposed model is enhanced by visualizing the output feature distribution of the refined time-frequency convolution kernel. Finally, the effectiveness, interpretability and generalization of the proposed method are verified by two real aero-engine operation cases.