Multi-GAT: 基于多度量衡构建图的故障诊断方法

Multi-GAT: A fault diagnosis method based on multi-metrics construction graphs

中文关键词  图卷积神经网络  故障诊断  图注意力机制  深度学习

英文关键词  graph convolutional neural networks  fault diagnosis  graph attention mechanism  deep learning

中文摘要

      基于图神经网络的故障诊断方法, 通常需要根据度量衡确定样本之间的相似性, 进而构建图的拓扑结构.然而, 根据单一度量衡可能无法准确衡量数据样本之间的相似性, 进而导致无法准确表征样本之间的关系. 因此, 选用不同的度量衡会极大地影响图神经网络的诊断性能. 为了解决通过单一度量衡无法准确表征数据样本之间相关性的问题, 本文提出了一种基于多度量衡构造图的故障诊断模型???Multi-GAT. 通过结合3种度量衡的计算结果,从而判断数据样本之间相关性的强弱. 本文改进了图注意力网络的评分函数, 使其能够依据样本之间相关性的强弱更准确地确定数据样本之间的相似性. 在本文基准数据集上的实验表明, Multi-GAT能够提升模型的诊断精度,且拥有较好的稳定性.

英文摘要

      Fault diagnosis methods based on graph neural networks usually require determining the correlation between samples based on a metric, which in turn constructs the topology of the graph. However, the correlation between data samples may not be accurately measured based on a single metric, which in turn may not accurately reflect the relationship between samples. Therefore, the choice of different metrics can greatly affect the diagnostic performance of graph neural networks. In order to solve the problem that the correlation between data samples cannot be accurately characterized by a single metric, a fault diagnosis model, the multi-metrics graph attention network (Multi-GAT), is proposed to construct graphs based on multiple metrics. The strength of correlation between data samples is determined by combining the results of the three metrics. The scoring function of the graph attention network is improved to determine the similarity between data samples more accurately based on the strength of correlation between the samples. Experiments on a benchmark dataset show that Multi-GAT is able to improve the diagnostic accuracy of the model and has good stability