融合多域特征深度挖掘的水电机组振动故障递进式诊断策略研究

The hydraulic turbine unit (HTU) belongs to the complicated nonlinear dynamic system with a developing tendency of giant-sized, synthesized and high specific-speed features. These characteristics make it a tough work for HTU in dynamic behavior description and fault diagnostic analysis aspects. To improve the fault diagnosis quality of HTU from multiple hierarchies, a progressive vibration fault diagnosis system for HTU with the characteristics of signal multi-scale precise description and more information after feature space reduction is built by taking the deep mining of multi-domain features of the vibration signals as the point-cut and breaking through the shackle of the traditional single-step diagnostic mode. First, by investigating the mode mixing elimination mechanism of ensemble empirical mode decomposition (EEMD) under small-magnitude and high-frequency white noise circumstances, a fast-speed EEMD scheme with the self-adaptive selection of substitution parameters has been proposed to implement the precise description of partial information of the signals. Subsequently, the lower-dimensional presentation of multi-information in nonlinear high-dimensional feature space has been implemented by establishing the sample adjacent domain relations and the weight edges determinative rules of the Pearson’s correlation index and deducing the principal component analysis (PCA) and locality preserving projections (LPP)-based spatial information integrated reservation model. On the basis of the aforementioned theories, the permutation entropy analysis-based elementary fault diagnosis of HTU is studied and a complicated probabilistic statistical analysis and random forest-based progressive multi-vibration fault diagnostic strategy is presented. The research findings will systematically perfect the fault diagnosis theorem of HTU and lay theoretical foundation for the improvement in the intelligent operating maintenance level of HTU in China.

水电机组属于复杂非线性耦合动力系统，且呈现大型化、综合化、高比转速发展态势，使得机组动力学行为描述以及故障诊断分析更为困难。为多层次系统提升机组故障诊断品质，以振动信号多域特征深度挖掘为切入点，突破传统单步诊断模式束缚，建立具有信号多尺度精确描述和特征空间多信息约简特点的水电机组振动故障递进式诊断体系。首先探究小幅值高频白噪声下集成经验模态分解的模态混叠消除机理，提出替代参数自适应选择的快速集成经验模态分解方法，实现信号局部信息的精确描述；构建皮尔逊相似指标的样本邻域关系和边权重确定规则，推求基于主成分分析与局部保持投影的空间信息综合保留模型，实现非线性高维特征空间多信息低维表征；基于此，研究排列熵分析的水电机组故障初步检测，提出基于概率统计分析与随机森林的复杂多元振动故障递进式诊断策略。研究成果将系统完善水电机组故障诊断理论，并为提高我国水电机组运维智能化水平提供重要理论依据。

水电机组作为水电站能源转换的关键设备，呈现出大型化、综合化、高速化、临界化和智能化发展趋势。该系统运行过程中面临的故障安全问题日益突出，为推进机组设备状态检修策略的工程应用，基于人工智能的故障诊断理论与方法得到了广泛关注与研究。本项目从水电机组振动信号多尺度精细描述、多域特征深度挖掘、故障特征迁移学习和故障诊断策略等方面展开研究。针对机组振动信号的强非线性和非平稳特性，采用先进时频分析方法实现原始信号的多尺度分解，为从不同角度挖掘信号蕴含的特征信息，进而构建不同域空间下的混合熵特征集。同时，为了降低冗余信息干扰和提高敏感信息比重，基于随机森林模型构建了水电机组振动信号特征优选模型，实现了信号特征重要度的量化评估。考虑到故障诊断样本数据的不充分性和机组运行过程中各工况出现故障的不确定性，基于联合分布适配方法实现了故障特征集的跨工况迁移学习，进而结合最邻近结点算法实现了变工况下机组复杂故障模式的有效诊断。最终，依据工程实际中故障预警的快速性需求，提出了基于概率统计分析和智能模式识别方法的机组故障递进式分步诊断策略。上述研究成果对完善我国水电机组故障诊断体系具有一定支撑作用，并为提高水电机组运维智能化水平提供了重要理论依据。