Intelligent fault diagnosis has become increasingly relevant in predictive maintenance for rotating machinery. With advancements in data transmission and sensor technology, measurement systems can now gather vast amounts of data from multiple sensors. These multi-sensor datasets are multivariate time series with significant Spatial–temporal correlation. Utilizing this correlation to achieve accurate diagnostics is a significant challenge. To fully leverage the Spatial–temporal correlations, especially the correlations among different sensors at various time steps, we propose a new, richly connected Spatial–temporal graph neural network for diagnosing faults in rotating machinery. This network primarily comprises two modules: graph construction and graph diffusion convolution with pooling. The graph construction module initially builds a richly connected graph that considers both the connections among sensors at the same time step and the connections between sensors across different time steps. Subsequently, we design a attenuation matrix that takes into account temporal distances to adjust the connection strengths between sensors based on their time separation. By applying Graph Diffusion Convolution (GDC) on the constructed graph, information can be propagated among nodes within a broader neighborhood, even capturing the interactions between nodes across multiple time steps. By combining GDC with pooling operations, temporal and spatial dependencies can be effectively captured to learn efficient representations. We evaluated the effectiveness of our approach through comparative experiments on three datasets, challenging various methods. The results demonstrate our method’s superior capability in integrating Spatial–temporal features thoroughly.

中文翻译：

---

用于多传感器信息融合的旋转机械故障诊断的丰富连接时空图神经网络


智能故障诊断在旋转机械的预测性维护中变得越来越重要。随着数据传输和传感器技术的进步，测量系统现在可以从多个传感器收集大量数据。这些多传感器数据集是具有显著空间-时间相关性的多变量时间序列。利用这种相关性来实现准确的诊断是一项重大挑战。为了充分利用空间-时间相关性，特别是不同传感器在不同时间步长之间的相关性，我们提出了一种新的、连接丰富的空间-时间图神经网络，用于诊断旋转机械中的故障。该网络主要包括两个模块：图构造和带池化的图扩散卷积。图构造模块最初构建一个连接丰富的图，该图既考虑同一时间步传感器之间的连接，也考虑不同时间步传感器之间的连接。随后，我们设计了一个考虑时间距离的衰减矩阵，以根据传感器的时间间隔调整传感器之间的连接强度。通过在构建的图形上应用图形扩散卷积 （GDC），信息可以在更广泛的邻域内的节点之间传播，甚至可以捕获跨多个时间步长的节点之间的交互。通过将 GDC 与池化操作相结合，可以有效地捕获时间和空间依赖关系以学习高效的表示。我们通过在三个数据集上进行比较实验来评估我们方法的有效性，挑战各种方法。结果表明，我们的方法在彻底整合时空特征方面具有卓越的能力。