A semi-supervised learning framework fusing multi-sensor visual features for ship propulsion shafting fault diagnosis

Propulsion shafting is a vital component of the ship power system, and its failure seriously affects navigation safety. Leveraging multi-sensor data for more reliable diagnostics is imperative. Transfer learning techniques have demonstrated considerable benefits in ship fault diagnosis, but in multi-sensor scenarios, extracting domain-invariant and discriminative features remains a challenge. To tackle this issue, this study develops a three-stage adversarial training framework for the first time. The framework effectively fuses the bearing housing vibration response information and shaft displacement information to discriminate the alignment status of the propulsion shafting. One-dimensional signals are transformed into two-dimensional images using the Hilbert symmetrized dot pattern to enhance image textures. The Vision Transformer architecture is employed to ensure the optimal integration of multi-sensor vision information. Subsequently, combining the data distributional discrepancy measurement theory and the pseudo-label strategy, a subdomain matching method is presented based on the progressive training paradigm. Meanwhile, a dynamic pseudo-label rectification strategy is designed to guide the primary network for more precise matching. The experimental results confirm the effectiveness of the proposed framework, which outperforms other leading methods. It offers a scalable solution for the intelligent operation and maintenance of propulsion shafting, with significant engineering application value.