Noise-robust underwater propeller fault diagnosis through dictionary learning and graph convolutional networks

Significant challenges are faced in underwater propeller fault diagnosis due to severe noise interference and the scarcity of labeled fault data in marine environments. In this paper, a novel semi-supervised approach is proposed that synergistically combines dictionary learning and graph convolutional networks (GCNs) to address these challenges. Frequency domain features are first extracted from hydrophone signals using the Fast Fourier Transform (FFT), and dictionary learning is then employed to generate noise-resistant sparse representations. Based on these representations, a graph structure is constructed using Pearson correlation coefficients and K-nearest neighbors, which is subsequently processed by a three-layer GCN that is trained in a semi-supervised manner with a pseudo-labeling strategy. Comprehensive experiments on three propeller types under various fault conditions were conducted, and exceptional effectiveness and robustness were demonstrated. It was found through experimental results that not only is excellent performance achieved under normal conditions, but stable diagnostic capability is also maintained in high-noise environments, thus significantly outperforming existing techniques.