Denoising graph attention wavelet network: an interpretable multi-sensor information fusion diagnostic method for rotating systems

Abstract

The strong noise and black-box nature of deep networks pose great challenges to the efficiency of utilizing multi-sensor data for fault diagnosis. To solve these issues, a denoising graph attention wavelet network (DGAWN) is proposed for multi-sensor information fusion fault diagnosis of rotating machinery. Considering the spatial relationship of multi-sensor measurement points, k-neighborhood graphs are first constructed to characterize the intrinsic association and topology of each sensor data. The graph attention (GAT) mechanism is introduced into the DGAWN to enhance the feature mapping relationship of multi-sensor graph nodes. Importantly, a denoising graph wavelet convolution (DGWConv) layer is developed by utilizing a graph wavelet operator followed by an adaptive thresholding denoising (ATD) module, thus obtaining interpretable graph wavelet filter responses and enhancing noise immunity. With the synergy of the attention mechanism, the DGAWN model matches the interpretable features with fault attributes under noise environments by utilizing the locality and sparsity of the graph wavelet kernel. Finally, two challenging rotating machinery fault datasets are used to validate the proposed DGAWN method under noise environments. Compared with several state-of-the-art models, experimental results indicate that DGAWN achieves the highest average diagnostic accuracies of 99.9% and 100% and superior noise resistance.