旋转机械是能源动力、航空航天等工业领域机械装备的主体。开展旋转机械故障诊断研究对于装备安全稳定运行具有重要意义。随着人工智能的兴起，旋转机械故障诊断智能化是未来发展趋势。针对现有相关研究中仍然存在的两个问题：1）诊断模型训练样本与模型部署场景中的测试样本满足独立同分布的假设；2）利用充足的故障样本学习高精度的诊断模型，本文以深度学习、迁移学习为基础，从多源信息融合、诊断模型泛化能力、参数迁移、领域自适应等角度逐步开展研究，旨在为解决工程实际中典型故障样本缺乏、数据分布差异等问题提供新的思路。针对旋转机械复杂系统中的多源时域序列特征融合问题，本文在传统端对端深度学习故障诊断框架基础上，提出了一种基于时空特征的深度学习方法。该方法主要包括多源时域序列时空特征提取和基于深度卷积神经网络的设备状态分类。所提方法有效提高了故障诊断能力和应对未知工况诊断任务的泛化能力。针对不同工业诊断任务中的数据分布差异致使智能诊断模型部署场景局限性的问题，提出了一种深度参数迁移的故障诊断方法。首先在辅助域中对深度模型进行预训练，然后将模型参数迁移至目标任务，最后利用少量目标域样本进行有监督微调。为了避免模型在不充足样本条件下过拟合，对比研究了三种典型的参数迁移策略，为旋转机械故障诊断任务参数迁移提供指导。基于此，进一步改进了传统深度模型结构，有效提高了目标域训练样本不足时的迁移诊断准确率。针对旋转机械监测数据主要以无标签形式存在而无法对智能模型进行监督训练的问题，提出了一种基于无监督领域自适应的迁移诊断方法。通过对辅助域与目标域的特征分布进行适配，从而保证从辅助域故障样本中学习到的诊断模型可以泛化到目标域任务。在传统的边缘分布适配的基础上引入条件分布适配，构建基于联合分布适配的深度迁移网络，进一步提高了特征分布适配的精度。在跨工况、跨故障模式的迁移诊断实验中进行方法验证。针对智能故障诊断中故障数据极其有限的问题，提出了一种少样本条件下的迁移诊断方法。通过利用目标域少样本的机械状态类别信息，将其与辅助域同类故障样本进行配对。利用多重对抗网络处理配对的数据，旨在有针对性进行领域自适应，保证精准的特征分布适配。在两个典型迁移故障诊断场景中，即跨工况迁移诊断与跨设备迁移诊断，进行方法验证。

Rotating machinery is widely used in industrial applications. With the trend towards more precise and more critical operating conditions, mechanical failures may easily occur. Condition monitoring and fault diagnosis is an effective tool to enhance the reliability and security of rotating machinery.Recent years have witnessed increasing popularity and development of artificial intelligence spanning through various fields. However, most of the existing studies have been performed with the two assumptions, which are not in accord with situations in real diagnosis tasks: 1) the training and testing data are taking from same distribution; 2) A large amount of training data is often required. To tackle these problems, a fault diagnosis framework based on deep transfer learning is presented. This study focuses on different aspects, including multi-source information fusion, the generalization capability of diagnostic model, parameter transfer, domain adaptation, so as to facilitate the application of intelligent fault diagnosis in real industry tasks.For the numerous measurements from diverse subsystems or components, the collected data is with disparate characteristics and needs more prevailing methods for data preprocessing, feature extraction and selection. This work presents a novel diagnosis framework that combines the spatiotemporal pattern network (STPN) approach with convolutional neural networks (CNN) to build a hybrid ST-CNN scheme. The proposed framework presents superior diagnostic accuracy and generalization ability, which is essential for the application in machine fault diagnosis. To improve the applicability and flexibility of diagnostic model across diverse tasks, this work proposes a diagnosis method based on deep learning and parameter transfer. First, the deep model is trained on large datasets to learn the hierarchical features from the raw data. Then, the parameters of the pre-trained model are transferred to new tasks with proper fine-tuning instead of training a network from scratch. Three transfer learning strategies are discussed and compared to investigate the applicability as well as the significance of feature transferability from the different levels of a deep structure. Moreover, a modified CNN is presented to improve the transfer performance with limited data.To utilize the large amount of unlaleled data in condition monitoring systems, a new intelligent fault diagnosis framework, i.e., deep transfer network (DTN), which generalizes deep learning model to unsuperived domain adaptation scenario, is proposed in this work. Domain adaptation is capable of minimizing the distribution discrepancy between domains, and thus the trained model in source domain can generalize well in target tasks. By extending the marginal distribution adaptation (MDA) to joint distribution adaptation (JDA), the proposed framework can guarantee a more accurate distribution matching. Extensive empirical evaluations on three fault datasets validate the applicability and practicability of DTN, while achieving many state-of-the-art transfer results in terms of cross-condition and cross- fault type.Consdering that only extremely limited fault data, namely sparse data (single or several samples), can be obtained, this work presents a novel framework for disposing the problem of transfer diagnosis with sparse target data. In consideration of the unclear data distribution described by the sparse data, the main idea is to pair the source and target data with the same machine condition and conduct individual domain adaptation so as to alleviate the lack of target data, diminish the distribution discrepancy as well as avoid negative transfer. The extensive experiments on two case studies, i.e., cross-condition and cross-equipement, are used to verify the proposed method.