A novel graph representation framework for bearing fault diagnosis via wavelet packet decomposition and Gramian angular field weighting

Traditional fault diagnosis methods often fail to capture multiscale vibration signal features and inter-band correlations effectively. To address this limitation, this paper proposes a novel framework combining Wavelet Packet Decomposition (WPD) and Gramian Angular Field (GAF) to transform vibration signals into structured graphs. WPD extracts multi-resolution features (energy, entropy) as graph nodes, while GAF quantifies temporal correlations between frequency bands as edge weights, preserving critical fault signatures. To mitigate reliance on large-scale labelled data, an adaptive data augmentation strategy is designed to prioritise important nodes, and a Graph Attention Network (GAT) with Graph Contrastive Learning (GCL) pre-train maximises feature separability with minimal labelled data. Two open bearing fault datasets and one bearing fault diagnosis experiment are applied to validate the performance of the proposed method. These experimental results demonstrated the superiority of the proposed method, achieving above 97% classification accuracy with small labelled training samples (67 per fault category), significantly outperforming other five current CL-based methods. In addition, the analysis of model parameters (GAT head and temperature coefficient) indicates that the proposed method can obtain the best diagnostic performance when GAT head and temperature coefficient be set to 8 and 0.5 respectively.