毕业生知识库

复杂工业设备的正常运行是保障制造过程安全稳定的前提，而构建实时感知并主动预测设备状态的健康管理系统是实现复杂工业设备正常运行的重要手段。数据驱动的智能故障诊断方法作为健康管理系统的关键技术，旨在利用先进的信息技术（如机器学习）从海量数据中学习特征与故障之间的映射关系，归纳出故障的表征和诊断规则。然而，实际工业场景中数据量大而不全，故障数据稀缺并与运行工况等因素高度相关，相同故障类别的数据呈现出碎片化和非同分布的特点。这导致基于同域数据的智能故障诊断方法在跨域故障诊断任务中泛化能力较弱，极大限制了该类方法的应用性能。为了突破上述的限制，本论文借鉴迁移学习的思想，利用其他来源的非同分布数据来辅助构建诊断模型，通过知识泛化的方式提升模型在新任务上的诊断性能，实现面向故障数据稀缺问题的跨域故障诊断。主要研究内容和创新成果归纳如下：

（1）研究面向未知领域的跨域故障诊断问题，提出了一种基于显式特征正则的元学习领域泛化方法EFRMAML。该方法利用双层优化模拟领域间数据分布偏移，通过领域间梯度匹配来学习多源域的共性知识；在梯度匹配的基础上利用显式特征正则方法约束特征学习；简化EFRMAML算法流程，通过混合域内层优化参数的直接更新提高算法执行效率。所提方法不需要利用目标域的故障信息，平均诊断准确率较基线模型提升了7.2%，达到了与利用目标域边缘概率分布信息的无监督领域自适应方法同等水平的诊断性能。

（2）研究面向不平衡领域的跨域故障诊断问题，提出了一种基于标签传播的元学习领域自适应方法LPMAML。该方法改进了自训练标签传播算法的优化目标，融合领域自适应方法实现对目标域的直接学习；利用双层优化模拟教师-学生模型训练过程，通过自训练和一致性正则实现源域监督信息的传播；利用基于采样方法的领域间隐式联合概率分布对齐方法减小标签传播方法的类别偏见。所提方法能够适配主流的领域自适应方法，使得决策边界更加鲁棒，显著提升原领域自适应方法的诊断性能，诊断准确率提升约8%，有效降低了类别不平衡造成的泛化误差风险。

（3）研究面向小样本领域的跨域故障诊断问题，提出了一种基于孪生架构的深度原型网络SADPM。该方法利用实例-原型相似度约束挖掘对比样例的全局领域原型来对齐领域间的语义关系，从而实现知识迁移；基于实例-原型的距离度量建立已知样本与领域原型的距离分布函数，自适应地调整首次故障的识别阈值，实现对未知故障的预警。所提方法在各个类别仅有极少量样本的条件下诊断准确率达到99.5%，显著提高了模型的诊断性能。

（4）集成上述三项关键技术，设计了跨域智能故障诊断方法的诊断流程，利用搭建的滚动轴承故障模拟试验台验证目标域信息未知、目标域具有无标记样本以及目标域具有少量有标记样本这三类目标域信息缺失条件下所提方法的跨域诊断性能。论文开发了基于工业大数据平台的智能故障诊断系统，用于诊断锂离子动力电池电极段关键设备的轻微故障，为故障数据稀缺场景中的跨域故障诊断任务提供技术支撑，提高了复杂工业设备运行的安全性和可靠性。应用结果表明，论文设计的智能故障诊断系统通过逐步积累故障知识有序提升模型的诊断性能，在各故障类别仅有5个样本的条件下轻微故障的诊断准确率超过90%，同时具有对未知故障的识别能力。

The normal operation of complex industrial equipment is a prerequisite for ensuring the safety and stability of the manufacturing process. Constructing health management systems that can timely perceive and proactively predict the health condition of the equipment is a significant means to achieve the normal operation of such equipment. Data-driven intelligent fault diagnosis methods, as the key technology of health management systems, aim to learn the mapping relationship between features and faults from massive data using advanced information technologies (such as machine learning) and induce fault characteristics and diagnosis rules. However, in actual industrial scenarios, the amount of data is large but incomplete, and fault data are scarce and highly correlated with operating conditions and other factors. The data of the same fault present fragmentation and non-identical distribution. This leads to weak generalization capabilities of the intelligent fault diagnosis methods based on same-domain data in cross-domain fault diagnosis tasks, greatly limiting the application performance of these methods. To break this limitation, this dissertation, inspired by the idea of transfer learning, uses non-identically distributed data from other sources to assist in constructing of diagnosis models. It promotes the diagnosis performance in new tasks through knowledge generalization with respect to the fault data scarcity problems of cross-domain fault diagnosis. The main research contents and innovative achievements are summarized as follows:

(1) The cross-domain fault diagnosis problem for unknown domains. An explicit feature regularization-based model agnostic meta-learning (EFRMAML)-based domain generalization method is proposed. This method utilizes a bi-level optimization strategy to simulate data distribution shift between domains and learns the common knowledge of multi-source domains by gradient matching. Based on gradient matching, an explicit feature regularization-based method is proposed to constrain feature learning. The EFRMAML is simplified by directly updating the parameters using the inner-layer optimization of the mixed domain to improve execution efficiency. Without fault information in the target domain, the proposed method achieves a diagnosis performance equivalent to unsupervised domain adaptation methods that use marginal probability distribution information in the target domain, with an average diagnostic accuracy improvement of 7.2% compared to the baseline model.

(2) The cross-domain fault diagnosis problem for class-imbalanced domains. A label propagation-based model agnostic meta-learning (LPMAML)-based domain adaptation method is proposed. This method improves the optimization objective of the self-training-based label propagation algorithm and integrates domain adaptation methods to achieve direct learning of the target domain. It uses a bi-level optimization strategy to simulate the teacher-student model training process and propagates the source domain's supervised information through self-training and consistency regularization. Then an implicit joint distribution alignment method is proposed based on a sampling method to reduce the class bias of the label propagation method. The proposed method can adapt to mainstream domain adaptation methods, making the decision boundary more robust. It significantly improves the diagnosis performance of the original domain adaptation methods, with an average diagnostic accuracy improvement of about 8%. Meanwhile, it effectively reduces the generalization error risk caused by class imbalance.

(3) The cross-domain fault diagnosis problems for small-sample domains. A Siamese architecture-based deep prototypical network (SADPM) is proposed. This method transfers fault knowledge by the instance-prototype similarity constraint, which mines global domain prototypes of contrastive samples to align semantic relationships between domains. Meanwhile, the distance distribution function between known samples and domain prototypes is established by the distance measure between the instance and prototype. It adaptively adjusts the recognition threshold for the first occurrence of faults, achieving early warning for unknown faults. The proposed method achieves a diagnostic accuracy of 99.5% under conditions where only a few samples are available for each class, significantly improving the model performance.

(4) Integrating the above three key technologies. The dissertation designs the diagnostic process of the cross-domain intelligent fault diagnosis methods and uses the built rolling bearing fault simulation (RBFS) test rig to verify the cross-domain diagnosis performance of the proposed method under three conditions of information deficiency in the target domain: target domain without information, target domain with unlabeled samples, and target domain with a few labeled samples. An intelligent fault diagnosis system based on an industrial big data platform was developed and applied to diagnose minor faults of the key equipment in the electrode manufacturing process of lithium-ion power battery manufacturing. It provides technical support for cross-domain fault diagnosis tasks in fault data scarcity scenarios and improves the safety and reliability of industrial equipment operation. Application results show that the designed intelligent fault diagnosis system gradually improves the model's diagnosis performance through the orderly accumulation of fault knowledge. It achieves a diagnostic accuracy of over 90% with only 5 samples per class and can recognize unknown faults.