On cross-attention-based graph neural networks for fault diagnosis using multi-sensor measurement

Rotating machinery fault diagnostics has received a lot of attention in recent years. As a result, there has been an increase in research interest in rotating machine intelligent fault detection, especially when using measurement from multi-sensors. However, an accurate fault diagnosis is still challenged based on nonlinear and non-stationary vibration signals. On the other hand, not enough research has been done on structural information fusion from multi-sensor measurement due to the complexity of spatial–temporal correlation. This paper explores the use of vibration signals in multi-sensor measurement fusion for rotating machinery fault diagnostics, and a method using a cross-attention-based dual-branch graph neural network (CA-GNN) is proposed. First, improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) is used to decompose vibration signals. Variational mode decomposition is then used to further deconstruct the highest-frequency subsequence after its parameters have been optimized using the whale optimization algorithm. Next, we build the CA-GNN, which contains two space branches for high- and low-frequency signals after initially creating the graph using multiple sensor data sets. Information across both branches for high- and low-frequency components can then be efficiently fused. Lastly, two experimental scenarios are used to illustrate the suggested technique and assess its viability and accuracy for rotating machinery fault diagnosis. Results indicate that the proposed method can diagnose rotating machinery health issues with an average accuracy of up to 99%, indicating that the method’s performance can fulfill real-world requirements.