基于迁移学习的旋转机械故障诊断及寿命预测方法研究

Constructing universal fault diagnosis and remaining useful life (RUL) prediction models is one of the most challenging tasks due to the inevitable indirect measurement problem of complex mechanical systems as well as their variable working conditions. This project first investigates the similarities and diversities existed in the mechanical system caused by varying working conditions and different types of faults. The fault-related features are also extracted from the three time-frequency and spatial domains. Then, transfer learning principles and its applications will be studied thoroughly, aiming to overcome the narrow focus of the conventional diagnostic approaches, from four different (i.e., instance-weights, feature distributions, associated rules and model parameters) aspects. Furthermore, a judgment criterion for transfer possibility will be built, and the shared preferences of training domains and target domains are extracted to help improve the classification accuracy. This project will also take advantages of the high-knowledge transformation ability of transfer learning to implement the universal RUL prediction of rotating machine components by building the dual-scale dynamical model to estimate the tracking indicators which reflect physical damage of the machine components. Both the fault diagnosis and RUL prediction models will be verified through experiments in the laboratory and real-world wind power drive systems. It is expected that the proposed project, if successful, will provide a solution for generic fault diagnosis and RUL prediction model construction. It will also advance the development of machine fault diagnosis and RUL prediction technologies.

复杂机械系统的工况不确定性以及局部部件信息的不可直接测量性，使得建立通用的故障诊断与剩余寿命预测模型成为当前颇具挑战性的研究方向。本课题首先研究机械系统在不同工况条件下、不同故障类型时所具有的共性以及差异性，并在时频空间域上提取故障特征。随后深入学习迁移学习理论，从基于实例权重、特征分布、关联规则以及模型参数几个方面开展研究以克服传统诊断方法针对面窄的问题，在此基础上建立可迁移度评判标准，提取训练与目标域的共享参数，以提升分类的准确度。同时利用迁移学习的知识转化能力，以双尺度动力学系统建模为手段，建立与物理损伤相关联的跟踪指标，开展旋转部件寿命预测研究，模型将在模拟试验台以及实际风电设备上进行测试与验证。可以预见，本课题提出的基于迁移学习的故障诊断与寿命预测研究可解决模型通用性问题，为分析具有多样性故障特征的复杂机械系统提供有价值的手段，有望推动机械故障诊断与寿命预测技术的发展。

复杂机械系统的工况不确定性以及局部部件信息的不可直接测量性，使得建立通用的故障诊断与剩余寿命预测模型成为当前颇具挑战性的研究方向。本课题首先研究机械系统在不同工况条件下、不同故障类型时所具有的共性以及差异性，并在时频空间域上提取故障特征。随后深入学习迁移学习理论，从基于实例权重、特征分布、关联规则以及模型参数几个方面开展研究以克服传统诊断方法针对面窄的问题，在此基础上建立可迁移度评判标准，提取训练与目标域的共享参数，以提升分类的准确度。利用迁移学习的知识转化能力，实施了基于奇异值分解/TrAdaboost的故障诊断研究；实施了基于谱质心迁移的故障诊断研究；实施了基于样本和特征混合迁移的机械故障诊断研究；实施了基于快速自组织映射迁移的机械故障诊断研究；实施了基于改进LSSVM迁移学习的故障诊断研究；实施了基于改进概率潜在语义分析模型应用于机械故障诊断研究；实施了基于时间迁移和空间迁移的机械故障诊断研究；实施了基于惩罚域选择机迁移的机械故障诊断研究；实施了基于深度迁移学习的机械故障诊断研究；实施了基于TCCHC的指数半随机EKF的机械剩余寿命预测研究。共发表期刊论文24篇、会议论文10篇，其中SCI收录15篇（包括ESI热点论文一篇、高被引论文4篇 ），两篇中文期刊论文获优秀论文奖，获国内会议优秀论文奖1次，国际会议最佳论文奖1次；申请发明专利5项，其中已授权4项；负责人入选ASME Fellow, 百千万工程人才国家级人选; 培养博士生3人（已毕业1人）、硕士生3人（已毕业2人）。本课题提出的基于迁移学习的故障诊断与寿命预测研究可解决模型通用性问题，为分析具有多样性故障特征的复杂机械系统提供有价值的手段，能够推动机械故障诊断与寿命预测技术的发展。