An interpretable multiscale graph wavelet neural network for aeroengine fault diagnosis under time-varying speeds

Graph neural networks (GNNs) have emerged as effective tools for aeroengine fault diagnosis due to their ability of fusing multi-sensor information. However, the multivariate time series from multiple sensors often exhibit stochastic spatio-temporal dynamics due to time-varying speeds, which poses a challenge in revealing intrinsic correlations within multi-sensor data and capturing fault representations. Besides, the lack of interpretability impedes the profound understanding of GNN-based fault diagnosis models. To address these issues, this paper proposes a multiscale spectral graph wavelet convolutional fusion network (MSGWCFN). Specifically, MSGWCFN constructs dynamic spatio-temporal graphs to reflect the complex dependencies between multi-sensor monitoring signals under time-varying speeds. To capture dynamic fault representations and enhance interpretability, a multiscale spectral graph wavelet convolutional (MSGWConv) layer is designed by integrating a set of spectral wavelet filters, which are derived from expert knowledge. Besides, a spectral kurtosis-guided readout (SKReadout) layer is developed to adaptively fuse the node representations of different sensors and output accurate diagnosis results. Experimental results on two challenging aeroengine fault datasets show that MSGWCFN achieves the highest diagnostic accuracies of 97.70% and 96.38%, respectively. Square envelope spectrum analysis shows that the learned multiscale filtering response can reveal characteristic frequencies that align with underlying fault mechanisms. These results suggest that the proposed MSGWCFN can bridge multi-sensor information and interpretable graph neural networks to play important roles in aeroengine fault diagnosis.