Fault diagnosis in rotating machinery using transfer learning combined with graph neural network algorithms with one-shot learning

Abstract In response to the issue of insufficient training data in the diagnosis of faults in rotating machinery, this research introduces a one-shot learning methodology that combines graph neural networks (GNNs) with transfer learning techniques. Rotating machinery vibration signals were processed using the short-time Fourier transform (STFT) to generate time-frequency spectra for the training dataset. A convolutional neural network (CNN) then extracted feature vectors and one-hot encoded representations from these fault diagrams. These extracted features subsequently served as inputs to the graph neural network (GNN) during training. The model’s architecture and parameters were meticulously defined, with overfitting effectively mitigated through dropout regularization. Finally, transfer learning was applied: parameters learned from source domain data were frozen, followed by fine-tuning the top layers using small-sample data under distinct operating conditions. Validation employed the publicly available Case Western Reserve University (CWRU) bearing dataset alongside data collected from a custom-built experimental setup. Comparisons with various one-shot learning methods further confirmed the superiority of the proposed framework. The results demonstrate improved performance on both experimental datasets.