We analyze vibration signals from wind turbines with dictionary learning and investigate the relation between dictionary distances and faults occurring in a wind turbine output shaft rolling element bearing and gearbox under different data and compute constraints. Dictionary learning is an unsupervised machine learning method for signal processing, which permits learning a set of signal-specific features that have been used to monitor the condition of rotating machines, including wind turbines. Dictionary distance is one such feature, and its effectiveness depends on an adequate selection of the dictionary learning hyperparameters and the data availability, which typically is constrained in condition monitoring systems for remotely located wind farms. Here we evaluate the characteristics of the dictionary distance feature under healthy and faulty conditions of the wind turbines using different options for the selection of the pretrained dictionary, the sparsity of the signal model which determines the compute requirements, and the interval between data samples. Furthermore, we compare the dictionary distance feature to the typical time-domain features used in condition monitoring. We find that the dictionary distance based feature of a faulty wind turbine deviates by a factor of two or more from the population distribution several weeks before the gearbox bearing fault was reported, using a data sampling interval as long as 24 h and a model sparsity as low as 2.5%.

我们通过字典学习分析来自风力涡轮机的振动信号，并研究字典距离与风力涡轮机输出轴滚动元件轴承和齿轮箱在不同数据下发生的故障之间的关系并计算约束。字典学习是一种用于信号处理的无监督机器学习方法，它允许学习一组信号特定功能，这些特征已用于监视旋转机器（包括风力涡轮机）的状况。字典距离是一个这样的特征，它的有效性取决于字典学习超参数的适当选择和数据的可用性，其通常在状态监测系统受限于远程定位的风力发电场。在这里，我们使用不同的选项来选择预训练词典，评估确定计算要求的信号模型的稀疏性以及数据样本之间的间隔，使用不同的选项来评估风力涡轮机健康和故障情况下词典距离特征的特性。此外，我们将字典距离特征与状态监视中使用的典型时域特征进行了比较。我们发现，在报告齿轮箱轴承故障之前几周，故障字典的基于字典距离的特征从人口分布中偏离了两倍或更多，使用的数据采样间隔长达24小时，模型稀疏度为低至2.5％。信号模型的稀疏性决定了计算要求，以及数据样本之间的间隔。此外，我们将字典距离特征与状态监视中使用的典型时域特征进行了比较。我们发现，在报告齿轮箱轴承故障之前几周，故障字典的基于字典距离的特征与人口分布有两倍或更多的差异，使用长达24小时的数据采样间隔和低至2.5％。信号模型的稀疏性决定了计算要求，以及数据样本之间的间隔。此外，我们将字典距离特征与状态监视中使用的典型时域特征进行了比较。我们发现，在报告齿轮箱轴承故障之前几周，故障字典的基于字典距离的特征与人口分布有两倍或更多的差异，使用长达24小时的数据采样间隔和低至2.5％。