旋转机械设备在实际工业生产中发挥着重要的作用，被广泛应用于电力、化工、矿业、航空等行业。对旋转机械设备开展故障诊断研究以保障设备安全稳定运行，避免意外停机所造成的经济损失和人员伤亡具有重大意义。基于数据驱动的智能故障诊断方法可以降低对设备故障机理和专家经验的依赖，得到了广泛的研究。现有研究大多依赖于训练数据充足、数据独立同分布、封闭世界假设等简单条件，而这在实际的复杂工业场景中可能难以满足。在此背景下，本文基于深度学习理论从知识迁移的角度出发，对实际工业场景中所存在的运行工况变化、有限故障数据及增量故障类型等复杂条件下的旋转机械设备故障诊断方法进行系列研究。本文的主要研究内容如下：针对变工况条件下的旋转机械故障诊断问题，本文提出了一种基于无源域自适应的跨域故障诊断方法，通过不同工况间的知识迁移实现跨工况故障诊断。首先在源工况下使用充足的标记数据对源故障诊断模型进行训练，然后使用目标工况下的无标签数据对源诊断模型进行基于伪标签聚类策略的自训练，实现源工况向目标工况的知识迁移并生成目标诊断模型，提升在目标工况下的故障诊断性能。在两个旋转机械数据集上的跨工况故障诊断实验验证了所提方法的有效性。针对小样本数据条件下的旋转机械故障诊断问题，本文依据元学习理论提出了一种基于不对称分布度量网络的小样本故障诊断方法。首先从具有充足标注数据和丰富故障类型的源域中构造大量的元诊断任务对小样本诊断模型进行元训练，使诊断模型学习任务无关的广义诊断知识。然后将小样本诊断模型迁移到新场景下的小样本故障任务中实现准确的故障诊断。所设计的方法在跨设备和跨故障类型的小样本故障诊断任务实验中取得了优秀的性能。针对不断出现的设备新故障类型数据，本文基于持续学习理论设计了一种类增量故障诊断方法，实现了诊断模型对新故障类型数据的增量学习和诊断模型的动态扩展更新。所设计的方法结合样本回放策略和知识蒸馏技术，通过新旧诊断模型间的知识蒸馏实现诊断知识的迁移和积累，同时引入权重对齐策略矫正分类器模型对新旧故类型的权重偏差，缓解诊断模型在新故障类型数据的学习过程中对旧故障类型知识的遗忘。在一系列的类增量故障诊断实验中对所提出的方法进行了验证，结果表明所提出的方法对新故障类型数据具有较好的增量学习能力。

Rotating mechanical equipment plays an essential role in practical industrial production. Conducting fault diagnosis research on rotating machinery equipment to ensure safe and stable operation, and to avoid economic losses and casualties caused by unexpected shutdowns, is of great significance. The data-driven intelligent fault diagnosis method for rotating machinery can reduce the dependence on equipment fault mechanisms and empirical knowledge and has been widely studied in recent years. Most existing research relies on simple conditions such as sufficient training data, independent and identically distributed data, and closed-world assumptions, which may be difficult to meet in practical complex industrial scenarios. In response to these issues, this thesis conducts a series of research on fault diagnosis methods for rotating machinery equipment under complex conditions such as changes in operating conditions, limited fault data, and incremental fault types in actual industrial scenarios from the perspective of knowledge transfer based on deep learning theory. The main research contents of this thesis have the following several respects:To solve the problem of fault diagnosis under variable operating conditions, this paper proposes a cross-domain fault diagnosis method based on source-free domain adaptation to transfer knowledge across operating conditions. Firstly, the source fault diagnosis model is trained using sufficient labeled data under source conditions. Then, the unlabeled data under the target condition is used to self-training the source diagnostic model based on a pseudo-label clustering strategy, achieving knowledge transfer from the source condition to the target condition, and generating the target diagnostic model, improving diagnostic performance under the target condition. The effectiveness of the proposed method is verified by experimental results of cross-working condition fault diagnosis on two rotating machinery datasets.Aiming at the problem of fault diagnosis with limited labeled fault data, this thesis designs a few-shot fault diagnosis method based on asymmetric distribution metric networks and meta-learning theory. Firstly, a large number of meta-diagnostic tasks are constructed from the source domain with sufficient labeled data and rich fault types for meta-training. The diagnostic model can learn task-independent generalized diagnostic knowledge in the meta-training stage. Then, the few-shot diagnosis model is migrated to the few-shot fault tasks in the new scenario to achieve accurate fault diagnosis. The designed method has achieved excellent performance in few-shot fault diagnosis experiments across different devices and fault types.In response to the continuous emergence of new fault-type data in rotating machinery equipment, this paper designs a novel class-incremental fault diagnosis method based on continual learning theory, which achieves incremental learning of the diagnostic model for new fault-type data and dynamic expansion and update of the diagnostic model. The designed method combines exemplar rehearsal strategy and knowledge distillation technology to achieve the transfer and accumulation of diagnostic knowledge through knowledge distillation between new and old diagnostic models. At the same time, a weight alignment strategy is introduced to correct the weight deviation of the classifier model for new and old types, alleviating the forgetting of old fault type knowledge by the diagnostic model in the learning process of new fault type data. The proposed method has been validated in a series of incremental fault diagnosis experiments, and the results show that it has good incremental learning ability for new fault-type data.