基于流形学习的风电系统传动部件多故障诊断及退化状态识别

The complexities of structure of equipment and assembly technology of transmission components decide the high probability of concurrent faults condition. In general, concurrent faults is multi-coupled, which is not simple linear superposition of single failure and causing difficulties to the fault diagnosis. Based on analysis of data attributes combined with correlation constraints, this work does research on transmission components of wind power mainly by approach of high-dimensional reconstruction, low-dimensional mapping, and isometric space conversion to decouple the fault features. As there are difference of the frequency band, energy and structure of the different types of fault signal, supervised manifold learning and multi-feature fusion are combined to extract fault feature, and the HSMM model that can identify degradation state of transmission components is also built. With the application of data-based fault feature extraction, degradation state recognition technology initiative failure prediction, achievement of this work can contribute to provide a theoretical and technical support for maintaining the operational reliability of the wind power system security and preventing failure, in condition of scarcity of existing wind power system fault data.

风电系统传动部件的复杂化使得多重并发故障发生的概率增大。多故障并发时，不同故障特征相互混杂呈现出多耦合的复杂征兆，并非多个单故障的简单线性叠加，给故障诊断造成困难。本项目以风电系统关键传动部件为研究对象，以数据结构分析为基础，通过加入去相关约束条件，将信号经历高维重构、低维映射以及等距空间转换，实现故障特征解耦。考虑不同种类故障信号的频带、能量、结构、形态等均不同的特点，采用多流形结构，联合多特征融合结果，经过有监督的多流形学习，实现特征提取，并在故障演化机理分析的基础之上构建HSMM模型辨识风电系统关键传动部件的性能退化演变状态。 本项目采用基于数据的故障特征提取、退化状态识别技术实现主动的故障预测，在现有风电系统故障数据资料匮乏的条件下，有望对保持风电系统安全可靠性运行，预防故障提供一定的理论及技术支撑。

风电作为重要的清洁能源，已经成为我国供电系统关注焦点。为了确保风电电力系统的供电稳定，必须保障大型风电系统的安全可靠运行。针对风电系统传动部件运行过程中出现的多故障情况，对由于故障混杂引起的特征耦合所带来的故障诊断难题进行研究。本项目以风电系统关键传动部件为研究对象，基于动力学分析，建立了传动部件的数学模型；基于数据挖掘思想，利用线性判别分析LDA进行故障数据降维处理，将反映故障的特征数据映射到特定的空间；采用有监督非相关邻域保持投影算法，以数据间的协方差值为约束条件，采用流形学习算法实现所提取的特征向量非相关性；以经验模态分解方法联合多重分形实现特征参数提取，经过流形学习实现特征向量降维；建立了基于ANN-HSMM网络的设备退化状态识别模型。这些成果有望对保持风电系统安全可靠性运行，预防故障提供一定的理论及技术支撑。