基于流形学习的机械设备早期故障诊断理论与关键技术研究

Aiming at characteristics of machinery early fault such as feature weak, feature coupling and dynamic development, a new idea and method based on manifold learning that posseses functions of multi-symptom domains feature fusion, decoupling, dynamic pattern recognition and newly added samples dynamic addition is proposed. Main contents include: ①researching on theoretical framework of “multi-symptom domains feature fusion →Feature compression with manifoldlearning→dynamic pattern recognition”, and constructing theoretical model based on this new mode for machinery early fault diagnosis; ②exploring new principle of weak fearture extacion based on manifold learning,and researching on structure and fusion Method of multi-symptom domains feature set, so as to comprehensive collecte and integrate early weark features of machinery; ③researching on decoupling method of strong coupling early fault feature,and introducing classification supervision mechanism and distance measure method, so as to effevtively decoupling and separate different early fault feature; ④ researching on dynamic pattern recognition model, and introducing mechanism of manifold space embedded intelligent decision, so as to dynamic track and accurate describe state variation of machinery; ⑤researching on mechanism of incremental learning, so as to solve newly added samples dynamic addition and adaptive learning and fusion of newly added samples and existing sample. This research project is not only greatly significant for improving level of condition monitoring and early fault diagnosis of machinery, but also highly valuable for ensuring safe, reliable and efficient running of the large-scale mechanisms.

针对机械设备早期故障征兆不明显、特征耦合、动态发展等特点，提出集多征兆特征融合、解耦、动态模式识别及新增样本动态添加等功能于一体的流形学习新思路、新方法。主要内容：①研究“多征兆域融合→特征约简→动态模式识别”的理论架构，构建基于新模式的机械设备早期故障诊断理论模型；②探索主流形学习的微弱特征提取新原理，研究多征兆域特征集构造和融合方法，实现对设备早期故障微弱特征的全面搜集和融汇；③研究强耦合早期故障特征解耦方法，引入分类监督机制和空间距离测度方法，实现对不同类型早期故障特征的有效解耦和分离；④进行动态模式识别模型研究，引入多流形空间嵌入智能决策和等距映射机制，实现对设备状态变化的动态跟踪和准确描述；⑤开展增量式学习机制研究，解决新增样本动态添加问题，实现新增样本与现有样本的自适应学习和融合。本项目的研究具有重要的理论意义，对于保障大型机械设备的安全、可靠、高效运行也具有重要的经济价值。

本项目针对机械设备早期故障的特点，开展基于流形学习的机械设备早期故障诊断理论及关键技术研究，目的是获得一种集多征兆特征融合、解耦、动态模式识别等功能于一体的流形学习方法。建立了耦合系统的非线性动力学模型，开展了旋转机械传动系统振动机理分析，研究了“多征兆域融合-特征约简-动态模式识别”理论架构，构建起了基于新模式的旋转机械早期故障监测诊断的理论模型；探索了多种微弱特征提取新方法，研究了多征兆域的特征集构造与融合方法；研究了内在维数估计和领域大小优化方法，实现了内在维数的最佳估计，获得了最佳的领域因子，构建了流形学习平台；研究了强耦合早期故障特征解耦方法，引入半监督判别机制，提高了故障分类能力；研发了一套旋转机械早期故障监测诊断系统。三年来，项目研究成果已在国内外公开发表学术论文8篇，录用待刊论文2篇，共计10篇，其中SCI检索2篇，EI检索5篇，录用待刊EI源期刊2篇，授权实用新型专利2项，依托项目正培养硕士研究生3名。项目负责人随着课题开展与深入不断成长，并入选浙江省151人才工程第三层次培养人选。此外，还进行了项目相关的扩展性研究，授权发明专利5项，实用新型专利3项。