针对船用发动机在变工况下故障数据匮乏导致的诊断难题，本文研究了更具挑战性的部分集迁移诊断场景，即目标工况的故障类别少于源工况。该场景下，源域中与目标域无关的故障类别会严重干扰诊断知识的迁移，导致现有深度学习方法性能显著下降。为此，本文提出了一种多尺度多视角域对抗网络（MMDAN）。该网络通过多尺度特征提取器捕获鲁棒故障特征，利用多视角分类器增强决策可靠性，并引入一个辅助域鉴别器来量化并筛选可迁移的源域样本，从而有效抑制了无关类别的负迁移效应。在某型船用发动机故障数据集上的实验表明，所提方法在多个变工况部分集任务中的平均诊断准确率达到96.58%，显著优于现有主流迁移学习模型，验证了其在实际复杂工况下的有效性与优越性。

The safe and stable operation of marine engines is critical to national security and maritime traffic safety. While numerous deep learning methods have been extensively studied for intelligent fault detection, the complex operating conditions of marine engines—such as non-stationary states including variable loads—often lead to prevalent domain shift problems in practical fault diagnosis tasks. This significantly degrades the performance of conventional deep learning approaches. Using a specific marine engine as a case study, we constructed partial-set fault diagnosis scenarios under varying operating conditions. To address the challenge of missing fault labels in the target operating condition, we propose a knowledge transfer approach from source to target operating conditions. A novel Multi-scale and Multi-view Domain Adversarial Network (MMDAN) is designed and experimentally validated using marine engine data. Experimental results demonstrate that the proposed method achieves an average diagnostic accuracy of 96.58%. Furthermore, in partial-set transfer tasks across different operating conditions, MMDAN exhibits superior diagnostic performance compared to other state-of-the-art learning models.