A fault diagnosis method for rotating machine based on feature fusion and heterogeneous graph neural networks

To address the challenge of multi-sensor information fusion in rotating machinery fault monitoring and diagnosis, this paper proposes a novel fault diagnosis method based on multi-sensor feature fusion using a heterogeneous neural network. First, after performing signal denoising preprocessing, both traditional knowledge-based features and deep learning features are integrated through an attention mechanism combined with KL divergence. This integration effectively reduces redundant fault information while enhancing the diversity and discriminative capability of the extracted features. Next, the fused features are fed into a heterogeneous neural network, in which multi-sensor signals are modeled as different types of nodes. A dynamically constructed adjacency matrix is employed to capture the complex correlations among multi-source data, thereby improving the collaborative representation ability between heterogeneous sensor features. Experiments conducted on a public bearing fault dataset and real-world motor data—covering both vibration and electrical signals—demonstrate the superiority of the proposed method. Compared with conventional CNN and CNN–GCN models, the proposed approach achieves higher accuracy, precision, recall and F1-score, validating its effectiveness and robustness in multi-sensor fault diagnosis tasks.