基于数据驱动的风电机组传动链故障诊断与性能预测研究

Large scale wind turbines work in poor environment and their laden changes are complicated, which lead to a high rate of mechanical fault ,a problem to evaluate the performance and status of components and so on.This topic focuses on the research of the drive chain fault diagnosis and performance prediction of wind turbines to realize regular health condition monitoring and performance analysis of wind turbines. A frequency spectrum analysis method based on equal angular sampling and a feature extraction algorithm based on local tangent space alignment are provided to solve the problem of fuzziness of spectral lines during the non-stationary process,the traditional Euclidean space algorithm can not find the essential information of the non-stationary signal. A condition monitoring and performance prediction model based on kernel entropy partial least square method(KEPLS) is proposed the first time, by introducing a new divergence measure statistics to achieve on-line monitoring, and the health condition of the transmission chain components as the dependent variable to achieve performance prediction. The forecast variance is used as index and the Bayesian inference is applied to calculate the confidence interval of the reliable prediction model. Study on a dual level partition network based on depth confidence network and affine propagation clustering,In order to solve the problem that the wind field is eager to make effective use of the historical fault data, the fault and grades of the existing fault vibration data of the wind farm are adaptively divided. Finally, the effectiveness and feasibility of the above methods are verified through the cooperation with wind power enterprises. The research is of great significance for the security and stable operation of wind power enterprises in the autonomous regions.

本课题针对大型风机工作环境恶劣，负载变化复杂，导致其传动链故障率高、部件性能与状态评估难等问题，研究风机传动链故障诊断与性能预测技术，实现对风机健康状态评估与性能分析。研究基于等角度重采样及局部切空间排列的特征提取算法，解决非平稳过程谱线模糊，传统欧式空间算法不能挖掘非平稳信号本质信息的难题。首次提出基于核熵排列的PLS(KEPLS)状态监测与性能预测模型，通过引入新的散度测度统计量实现故障在线监测，并将传动链部件健康状态作为因变量实现性能预估。其中以预测方差为指标，采用贝叶斯推理计算具有可靠性评价的预测模型置信区间。研究基于深度置信网络与仿射传播聚类的双层划分网络，对风场现有故障振动大数据的故障类型和等级进行自适应划分，解决风场迫切希望对历史故障数据进行有效利用的问题。最后通过与风电企业合作，将上述方法在风机上验证有效性及可行性。课题研究对于保障自治区风电企业安全、稳定运行具有重要意义。

当前针对风电机组传动链的故障诊断与性能预测存在如下难题：风电机组工作环境恶劣，负载变化复杂，导致其故障率高、性能评估难；风电机组常发生多种故障并存的复合故障，使得诊断难度更大；现场获取运行海量数据呈现非平稳、非线性特性，且是不完备和无标签的，亟需开发有效数据清洗和具有自学习能力的诊断评估算法。针对上述问题，本课题提出一套完整的故障诊断与性能预测新策略。具体成果包括：.(1)提出基于经验Copula-互信息(TFDD)的数据在线清洗方法，并利用维度规约与局部切空间排列对数据进行降维和特征提取。.(2)提出基于核熵排列的PLS(KEPLS)状态监测与性能预测模型，通过引入新的统计量实现故障在线监测，并将传动链部件健康状态作为因变量实现性能预估。研究过程分为三阶段递进实现，分别是构建KECA白化、在核熵空间建立PLS监控模型、通过KEPLS算法实现传动链部件健康状态预测。最终，形成一套针对高维、非线性、参量多的风电机组性能预测算法。.(3)提出一种基于孪生结构的双层网络框架实现风机传动链故障类型自学习和损伤等级自动识别。该研究将孪生结构与深度网络相结合，构建双通道故障自学习模型；之后将提取特征输入贝叶斯分类器，依据构建的后验概率判别规则实现损伤等级自动识别。该方法具有较好的自学习能力，增强了诊断的智能性。.(4)提出基于改进解卷积的风机复合故障诊断算法。针对风电机组发生复合故障的概率高，诊断难度大问题，将传统MOMEDA、CYCBD等解卷积方法进行优化改进，并成功应用于齿轮箱、滚动轴承等复合故障诊断中，取得较好的识别效果，具有一定的推广价值。.(5)延伸研究。课题组进一步研究了基于数学形态学分形维数的数据滤波算法，并结合特征频率知识实现故障定位。将上述方法进行融合，实现了从数据清洗、特征提取、状态监测、故障诊断和性能评估的全过程管理，课题的研究对于保障风电企业安全稳定运行具有重要意义。