Attention-based multi-modal learning for aircraft engine fan fault diagnosis

This research proposes a methodology using the deep learning framework to investigate rotating fan fault diagnosis - an endeavor of both scientific significance and practical importance. The methodology includes two critical steps: (1) the generation of high-frequency coupled acoustic mode spectrogram features and (2) the application of the attention mechanism-based feature fusion technique to integrate vibration Mel spectrograms, acoustic Mel spectrograms, and coupled acoustic mode spectrograms. The coupled mode spectrogram feature is adopted in deep learning research on fan fault diagnosis for the first time. The theoretical derivation of the coupled acoustic source , based on nonlinear acoustics , is provided for a deepened physical understanding of the proposed feature. Our model is verified and validated by analyzing its predictive performance on the representative datasets from our rotating fan experiments. According to the results, adding the coupled mode spectrogram feature can increase the model accuracy by at least 19% on small datasets. Furthermore, the accuracy of the attention-based model can be, at most, 13% greater than that of the straightforward feature fusion-based model. It reveals that the proposed classification model holds the potential for advancing aircraft engine fault diagnosis techniques and enhancing aircraft operational safety. • Innovative multi-modal learning for fan fault diagnosis. • First introduction of high-frequency coupled mode spectrogram. • Superior performance of attention-based fusion. • Enhanced model Accuracy in varying scenarios. • Successful fusion of the acoustic and vibration signals.