航空发动机作为大部分飞行载体的主要动力源泉，保障其安全可靠的运行对飞行安全至关重要。鉴于发动机长期处于严苛的工作环境下，且故障诊断是实现视情维修策略的关键技术，使得发动机故障诊断研究已成为航空动力系统领域的研究热点之一。为此，本文开展了针对气路的数据驱动航空发动机多故障诊断方法研究，针对发动机故障诊断所需经常面对的不同应用场景，分别制定了相应的应对策略。论文的主要研究内容及贡献如下： （1）研究了基于单分类算法在故障样本稀缺场景下的发动机气路故障检测技术，主要针对发动机运营初期故障数据难以获取的问题。利用一类模式识别方法，实现在仅有正常样本可利用的条件下，完成对发动机关键气路部件的故障检测任务。并提出快速约简一类极限学习机（FROC-ELM）算法，利用随机选取训练样本的策略来构建约简核函数矩阵，以解决借助核技巧的一类极限学习机（OC-ELM）算法存在的网络结构冗余的问题，从而实现在能够保持原有故障检测准确率的前提下，极大的降低算法在故障检测过程中训练与测试阶段的时间成本，提升检测模型的实时性能。 （2）针对故障诊断系统建立初期仅有少量故障数据可利用的问题，实施了类别不平衡场景下的发动机气路故障诊断研究，旨在解决训练数据集中正常数据与故障数据相比占比过高所带来的问题。为了缓解诊断模型可能对正常样本的过拟合及对故障样本欠拟合，提出了一种双重采样DS策略的诊断方法，通过快速消除多余样本和人工插值的策略，以缓解训练数据集中样本的类别不平衡程度。在不同类别不平衡程度下的发动机气路故障诊断结果表明，DS方法可在极端类别不平衡的情况下有效的提升诊断模型的分类性能。 （3）鉴于核极限学习机KELM算法使用所有训练样本来构造隐含层节点，将难以避免地导致网络拓扑结构冗余的缺点，提出一种群体约简核极限学习机（GRKELM）算法，利用稀疏学习中的 范数来实现输出权重中的群体稀疏结构，为选择最重要的隐含层节点创造条件。与此同时，以交替迭代优化的方式设计了一种高效快速的优化算法来解决GRKELM模型中存在的非平滑性优化问题，并给出了优化过程的收敛性分析及证明。通过基准数据集和发动机故障诊断试验，验证了GRKELM算法的诊断性能，使其可实现在不降低模型诊断精度的前提下，提升气路故障诊断系统测试阶段的实时性能。 （4）针对传统数据驱动的诊断方法通常存在过于理想化假设的问题，主要表现在训练数据（来自源域）和测试数据集（来自目标域）服从相同分布，开展了基于迁移学习的发动机气路故障诊断方法研究，并详细介绍了迁移学习中的一些基本概念。通过结合域自适应思想和极限学习机算法，提出了DTELM和JDTELM两种算法，使其在保留原始ELM网络拓扑结构简单这一主要优点的前提下，可通过学习可迁移的数据特征表示，减少源域和目标域数据之间分布差异，同时尽可能保留源域的数据属性和特征结构。其中，DTELM算法主要应对仅考虑边缘分布差异的情况，而JDTELM算法则使用经重构的源域数据来训练分类器，迭代的学习目标域中数据的伪标签信息，从而试图同时降低边缘分布和条件分布差异。并在此基础上，进行了大量的发动机气路故障诊断试验，同时诊断结果验证了算法的有效性。 （5）针对发动机气路并发故障诊断问题，开展了基于多标签学习的诊断方法研究。根据多标签学习中的问题转换策略，提出了MLFS-BR算法，利用稀疏正则化和标签相关性估计来建立前期的特征选择模型。当获取标签特定的特征后，可将其直接用于并发故障诊断任务。在面对MLFS模型中的非光滑凸优化问题时，采用了加速近端梯度法，并证明了优化模型中的Lipschitz连续性。最终通过一系列测试验证了MLFS-BR算法能够解决并发故障诊断问题。

Aero-engine is the main power source of most flight carriers, so it is very important for flight safety to ensure its safe and reliable operation. As the engine has always been working in a harsh environment, the research on engine fault diagnosis has become a research hotspot in the field of aerospace power system, and fault diagnosis is a key technical means to realize engine condition-based maintenance. To this end, this thesis has carried out aero-engine gas path fault diagnosis research based on data-driven diagnosis technology, and formulated corresponding response strategies according to the different application scenarios that the engine fault diagnosis needs to face frequently. The main research contents and contributions of the thesis are as follows: (1) In order to solve the problem that it is difficult to obtain fault data in the early stage of engine operation, the study on engine gas path fault detection technology based on one class classification algorithm is conducted. By this pattern recognition method, it is possible to complete the fault detection task of the key gas path components of the engine under the condition that only normal samples are available. Then a fast reduction one class extreme learning machine (FROC-ELM) algorithm is proposed, which can use the strategy of randomly selecting training samples to construct a reduced kernel function matrix, thus solving the problem of network topology redundancy in original kernel OC-ELM algorithm. Then, under the premise of maintaining the original fault detection accuracy, the time cost of the proposed algorithm in the training and testing phase is greatly reduced, and the real-time performance of the detection model is improved. (2) Given that a small amount of fault data is available in the initial stage of the fault diagnosis system, a research on engine gas path fault diagnosis addressing class imbalanced issues is implemented, aiming to solve the problem caused by the excessive proportion of normal data in the training dataset. In order to alleviate the over-fitting for the normal samples and the under-fitting for the faulty samples in the diagnostic model, a dual sampling (DS) strategy diagnosis method is proposed. Through the rapid elimination of redundant samples and manual interpolation strategies, the degree of class imbalance in the training data set is alleviated. Engine gas path fault diagnosis experiments show that the DS method can effectively improve the classification performance of the diagnostic model in the case of extreme class imbalance. (3) Given that the kernel extreme learning machine (KELM) algorithm uses all training samples to construct hidden layer nodes, it will inevitably lead to the shortcomings of network topology redundancy. To this end, a group reduction KELM (GRKELM) algorithm is proposed, which uses sparse learning and the -norm to achieve the group sparse structure in the output weights, and it is good the selection of the most important hidden layer nodes. Meanwhile, an efficient and fast optimization algorithm is designed to solve the non-smooth optimization problem in the GRKELM model by means of alternate iterative optimization, and the convergence analysis and complete proof of the optimization algorithm are given. Through a large number of experiments, including benchmark datasets experiments and engine fault diagnosis experiments, the feasibility of the proposed GRKELM is verified, so as to improve the real-time performance of the system test phase without affecting the accuracy of model diagnosis. (4) Traditional data-driven fault diagnosis methods always assume that the training dataset (source domain data) and test dataset (target domain data) follow the same distribution, which is too idealistic. To this end, the gas path fault diagnosis based on transfer learning is carried out, and some basic concepts are given. By combining the domain adaptation strategy and ELM algorithm, two algorithms, including DTELM and JDTELM, are proposed, which can algin the distribution discrepancy between source domain and target domain by learning transferable feature representations, while retaining the main advantages of the original ELM and preserving the data attributes and structure of the source domain as much as possible. DTELM algorithm mainly deals with the situation where only marginal distribution discrepancy are considered, while the JDTELM algorithm uses reconstructed source domain data to train the classifier and iteratively learns the pseudo-label information of the target domain data, thereby trying to align the marginal and conditional distribution discrepancies simultaneously. Then, a lot of experimental verification work has been carried out. (5) Aiming at the problem of engine gas path simultaneous fault diagnosis, a diagnosis method based on multi-label learning is implemented. According to the problem conversion strategy in multi-label learning, the MLFS-BR algorithm is proposed, which uses sparse regularization and label correlation estimation to establish the early feature selection model. After obtaining label-specific features, it can be used for multi-label classification tasks. When facing the non-smooth convex optimization problem in the MLFS model, the accelerated proximal gradient method is adopted, and the Lipschitz continuity in the optimization model is proved. Finally, experiments verify that MLFS-BR algorithm can handle simultaneous fault diagnosis problems.