变工况下旋转机械健康状态识别的深度迁移学习方法研究

The health status recognition of key components such as bearing is the key to ensure the stable operation of mechanical equipment during service. It is hard to pre-process the key component signals and extract the features under variable working conditions, especially when the early faults occur with weak characteristics. In addition, the equipment health status recognition model established by the traditional machine learning method is greatly affected by changes of working conditions and the universality of the model needs to be improved. To tackle the problem, this project intends to conduct the research on rotating machinery dynamic signal deep transfer learning model and recognize the mechanical health status under variable working conditions. Firstly, the sparse filter network and sparse criterion are constructed and optimized to explore the inherent structural characteristics of the mechanical status signal in the way of unsupervised learning. The features are expected to be multi-level enhanced effectively by removing the interference components and optimizing the computational efficiency. Then, due to the fact that the distribution of equipment status features is not consistent under variable working conditions and it is difficult to build a universal health status identification model, the transferable criterion is established based on the analysis of the deep relationship between conditions, objects and faults to ensure positive transfer from the perspective of fault mechanism. Finally, the deep knowledge transfer learning between training tasks is carried out to achieve effective mechanical health status recognition through data set optimization, dynamic reconfiguration, etc. The mechanical dynamic signal deep transfer learning model with better versatility under variable working conditions is developed, which provides a new way for ensuring the stable operation of mechanical equipment.

轴承等关键部件的健康状态识别是确保机械装备安全服役的关键。变工况下旋转机械关键部件状态特征，尤其是故障早期信号特征较微弱导致预处理困难；另一方面，传统机器学习方法建立的健康状态识别模型受工况的变化影响较大，模型通用性有待提升。本项目拟研究旋转机械动态信号深度迁移学习方法，开展变工况机械健康状态识别研究。首先，构建并优化稀疏滤波网络和稀疏准则，以无监督学习方式，探索变工况机械动态信号的内在结构特点，去除干扰成分，并进行算法时效优化，有效地多层次增强特征；然后，针对变工况下，机械状态特征分布不一致难以构建通用的识别模型问题，从故障机理角度，在分析工况、对象、故障之间深层次关系基础上建立迁移准则，确保模型正迁移；最后，通过优化数据集并进行动态重构等手段，进行训练任务间的知识深度迁移，实现机械健康状态的有效识别。研究通用性更强的机械健康状态深度迁移学习模型，为保障机械装备可靠运行提供一种新途径。

轴承等关键部件的健康状态识别是确保机械装备安全服役的关键。变工况下旋转机械关键部件状态特征，尤其是故障早期信号特征较微弱导致预处理困难；另一方面，传统机器学习方法建立的健康状态识别模型受工况的变化影响较大，模型通用性有待提升。面对这些挑战，本项目组分别从域适应和域泛化角度，研究通用性更强的机械健康状态深度迁移学习模型，本项目具体开展了以下几个主要方面的研究工作：1）提出了一种基于知识映射的对抗域自适应机械故障诊断方法，该方法能够通过领域鉴别器和特征提取器的对抗训练，将知识从目标域映射到源域；2）构建了类内多故障数据源域适应网络，该模型使用一个特征学习器来生成每个源域和目标域数据的特征，以使联合权重分类器能够预测目标标签。它还利用了一个基于矩量匹配的多源域距离度量准则来减少所有源域和目标域之间的距离。在模型的训练过程中，采用了类内对齐训练策略来同时对齐每个域的边缘和条件分布。单源域迁移基础上，充分利用了对模型有帮助的其他故障数据源域；3）提出一种基于类边界特征检测的对抗域泛化诊断网络，该模型可以在未知的工况下进行故障诊断，并且在训练中只使用一个完全标记域，有效避免了因标记数据不足引起的诊断精度不理想；4）提出了一种域不变特征学习域泛化学习基本框架，建立域泛化诊断模型，减少对目标域数据的依赖性，即如何通过多个可用的源域数据来学习未见过的目标域的广义故障特征表示。研究通用性更强的机械健康状态深度迁移学习模型，为保障机械装备可靠运行提供一种新途径。在本项目的支持下，本项目组发表论文51篇，其中SCI论文39篇，EI收录8篇，参加国际学术会议4次，国内学术会议8次；申请发明专利9项，其中5项已授权，申请并授权实用新型专利2项；培养博士研究生1名，硕士研究生9名。