Noise-robust fault diagnosis for rotating equipment in marine vessels using a band-selective envelope attention network without noisy data training

Abstract This paper addresses the problem of performance degradation that occurs under low signal-to-noise ratio (SNR) conditions when diagnosing rotating machinery faults in marine environments. Artificial intelligence-based rotating-machinery fault diagnosis exhibits performance limitations owing to restricted data availability in real environments and low-SNR conditions. Previous studies either trained models on noise-free laboratory data or incorporated noise-reduction modules during preprocessing to mitigate this problem. However, such approaches remain limited by a lack of prior knowledge of noise characteristics and amplitude, as well as poor generalization to unfamiliar noise conditions. To overcome these limitations, we propose a band-selective envelope attention network (B-SEAN). The proposed model divides raw vibration signals into multiple frequency bands using a band-splitting scheme and constructs stable envelope spectra for each band using the Hilbert transform and Welch’s method. The band attention module automatically downweighs bands that are severely distorted by noise while emphasizing those that preserve fault-related features. More importantly, the model parameters are optimized using only clean signals without injecting any noisy data during training. Using data from normal and seven fault conditions of a marine fuel-supply pump measured in a laboratory environment, we performed verification by synthesizing white Gaussian noise and Gaussian impulse noise in an SNR range of 5 to −6 dB. B-SEAN maintained superior performance under low-SNR conditions compared with existing noise-trained models. This demonstrates a practical method that enables reliable diagnosis despite strong noise in real operating environments by detecting nonlinear variations in fault signals and highlighting the fault features. In conclusion, B-SEAN satisfies both data efficiency requirements (as it can be trained without noisy data) and high robustness (as it maintains performance under various noise conditions). This study demonstrates the practical potential of noise-resilient fault diagnosis not only in marine vessels but also in a wide range of rotating machinery.