基于微调策略的跨工况船用柴油机故障诊断通用迁移学习框架

Abstract Timely and accurate fault diagnosis (FD) of marine diesel engines (MDEs) is crucial for enhancing the safety and reliability of ship power systems . MDEs operate under variable conditions, leading to significant differences in their operational and fault data. This variability reduces the adaptability of data-driven FD models, which are developed using data from a single engine or specific conditions. To address the aforementioned issues, this study proposes a fault diagnosis framework for MDEs based on deep transfer learning and fine-tuning. To enhance the capability of fault feature extraction, a data-tiered fusion method is introduced for data reconstruction. Furthermore, a novel hybrid pre-training network combined CNN + GRU and KAN is proposed to obtain comprehensive source domain data features. Additionally, a combined fine-tuning strategy for the transfer of pre-trained models is presented to enable superior cross-working-condition fault knowledge sharing. Experiments validate the effectiveness of the framework using two typical MDEs fault datasets. The proposed framework achieves 95 % accuracy in the MDE cross-working-condition (CWC) FD task, which is significantly better than the non-transfer model. The proposed framework not only demonstrates excellent performance in CWC fault diagnosis but also provides new insights for the paradigm of MDEs fault diagnosis.