Deep learning framework with dynamic oscillatory spatial attention for thrust bearing fault diagnosis in podded propulsors

The thrust bearing in podded propulsors represents a critical component of marine propulsion systems, for which precise fault diagnosis is vital to operational safety. However, accurate diagnosis under complex marine operating conditions remains particularly challenging, due to insufficient extraction of subtle fault characteristics and performance degradation in data-scarce scenarios. Therefore, an enhanced one-dimensional convolutional neural network was developed in this study to address these issues, i.e. the DOSART-1DCNN, in which multiple specialised modules are incorporated within an end-to-end deep learning architecture. The conventional spatial attention mechanism is refined in this study through the integration of a dynamic oscillatory coupling strategy. This employs adaptive oscillatory modulation, adjusting attention weights via a learnable oscillation function to form a Dynamic Oscillatory Spatial Attention (DOSA) module. Oscillatory attention is then combined with convolutional operations to process preliminary features and capture locally significant information. Then, lightweight residual blocks are applied to further refine feature representation, enhancing depth and alleviating gradient vanishing. And then, a Transformer module augmented with a dynamic sparsification factor improves local temporal modelling, enabling the capture of long-range dependencies and the integration of both local and global features. Consequently, the DOSART-1DCNN improves the detection of weak fault signatures while maintaining stability in deep architectures by embedding these components into a convolutional network backbone. Finally, the DOSART-1DCNN developed in this study is evaluated using a proprietary dataset, and the results showed a classification accuracy of 97.4 %. Additional validation was conducted using the JNU dataset to further demonstrate the superiority of the proposed DOSART-1DCNN. These results validate the efficacy of DOSART-1DCNN across multiple diagnostic tasks, highlighting its potential for reliable fault diagnosis in real-world marine propulsion applications.