Abstract: The scarcity of labeled fault sample data of wind turbine gearboxes in some wind farms seriously reduced the accuracy of fault diagnosis. To solve this issue, a fault diagnosis method based on mixed self-attention prototype networks under small samples was proposed. First, the vibration signals were mapped to the fault feature measurement space through the prototype networks. Then, the position self-attention mechanism and channel self-attention mechanism were used for matrix fusion to construct a mixed self-attention module, which established the global dependence of the original vibration signals and obtained more discriminative characteristic information to learn the measurement prototypes of wind power gearboxes in various health states. Finally, the trained metric classifier was adopted to identify the faults of the wind turbine gearbox under the condition of small samples. Experimental results show that the fault diagnosis method of the mixed self-attention prototype networks may achieve high-precision fault diagnosis of wind turbine gearboxes on different scales of small sample datasets.

Key words: deep learning, fault diagnosis, small sample, prototype network, mixed self-attention mechanism

摘要： 针对部分风场因有标签故障样本数据稀少而导致风电齿轮箱故障诊断准确率不高的问题，提出了一种小样本下混合自注意力原型网络的故障诊断方法。首先，通过原型网络将振动信号映射到故障特征度量空间;然后采用位置自注意力机制和通道自注意力机制进行矩阵融合构建混合自注意力模块，建立原始振动信号的全局依赖关系，获取更具判别性的特征信息，学习风电齿轮箱各健康状态下的度量原型；最后通过度量分类器进行模式识别，实现小样本条件下风电齿轮箱的故障诊断。实验结果表明，所提出的混合自注意力原型网络故障诊断方法在不同小样本数据集上均能实现风电齿轮箱高精度故障诊断。

关键词: 深度学习, 故障诊断, 小样本, 原型网络, 混合自注意力机制