风电并网变流器故障特征提取及诊断技术研究

It is a great challenge of diagnose the fault of grid-connected inverter in wind power generation systems. It because that the fault signal interaction and high voltage interference have great effect to weak fault signal detection. In this project, a weak signal detection method is employed to improve the weak signal detection performance under strong noise background. A novel method of the precise representation of fault time-frequency domain features is explored. This method could satisfy the requirement of the weak and potential fault feature extraction, fault source recognition and fault localization. Premeasure and integrated method based blind source separation is employed, the signal is separated into multiple components. Using the structured noise analysis modeling and sparse signal reconstructing method, the weak and potential fault signal from noise background is extracted. Using the extracted or separated signal to time-frequency space take the spectrum rotation transformation, the fault signal feature evaluation law from the time domain to frequency domain is analyzed. The fault feature is represented completely. Based on the fault information distribution manifold hypothesis, using the semi-supervised machine learning method taking the advantage of the massive unlabeled information to mining the unknown fault pattern then get the precise fault classify model. Finally, the grid-connected inverter fault is recognized and located.

本项目针对风力发电中并网变流器故障信号相互影响及强电干扰对微弱故障信号的影响，使得故障难以检测的问题，以提高强噪声背景下信号检测能力为切入点，探索故障信号的时频域特征的精细刻画方法，使之满足变流器微弱/潜在故障征兆的提取和故障源识别、定位的需求。通过基于预测度和集成学的盲源分离等方法，将信号分离成多个成分，并采用结构化噪声分析建模与稀疏信号重构的方式，提取出被噪声淹没的微弱/潜在故障信号；对分离或提取得到的信号，在时频域空间进行频谱旋转变换，分析故障信号从时域到频域的特征演变规律，更为完整地实现故障特征信息的描述；基于故障信息分布的流形假设，利用半监督机器学习方法，充分利用大量的未标注信息，挖掘未认知的故障模式，获取精确的故障分类模型；并基于证据理论对冗余信息和互补信息进行融合，解决故障诊断决策中系统检测信号量测的不确定性及故障征兆的不确定性，实现对变流器故障的识别和定位。

风电并网变流器中电力电子元件故障，直接影响风电系统风能转换与运行安全。本项目针对并网变流器故障征兆微小、故障种类繁多等问题展开了研究。首先对国内外并网变流器故障诊断研究现状进行了总结，利用Matlab等仿真软件，对并网变流器故障进行建模分析。结合现状总结以及定性分析结果，提出稀疏信号重构、线性正则变换等更精细的时频分析算法，实现并网变流器电信号微小特征提取。针对现有并网变流器故障库的不完备问题，结合电压电流等电信号的周期性，提出将迭代学习良好的模型逼近能力用于挖掘未标记的故障模式。针对并网变流器故障种类繁多且复杂的特点，将改进的聚类分析等智能算法用于并网变流器复合故障识别与定位。此外，为验证所提算法的有效性，搭建了风电并网变流器故障模拟实验平台。该验证平台具有并网控制、故障注入、多通道高速数据采集等功能，为研究并网变流器实际运行中的故障特性发挥了重要作用。