A multi-channel information fusion with adaptive weighting network for cross-domain fault diagnosis of rotating machinery

Abstract

In complex industrial environments, single monitoring signals, limited labeled data, and varying operating conditions often lead to low accuracy and poor generalization in cross-domain fault diagnosis of rotating machinery. To address these issues, a multi-channel information fusion with adaptive weighting network (MIFAWN) is proposed in this paper. First, the synchrosqueezed short-time Fourier transform (SSTFT) is employed to extract time-frequency features from axial, vertical and lateral vibration signals collected by triaxial accelerometers, which are then concatenated along the channel dimension to form fused time-frequency images. Second, a residual network integrated with an efficient channel attention (ECA) mechanism is constructed to dynamically focus on key features in directional channels. Furthermore, by combining dual-layer multi-kernel maximum mean discrepancy (MK-MMD) feature alignment with a domain adversarial training strategy, distribution adaptation between the source and target domains is achieved, enhancing the transferability and discriminability of features. To demonstrate the effectiveness of the proposed method, two types of transfer diagnosis tasks are carried out: cross-condition gearbox diagnosis and cross-device bearing diagnosis. Experimental results show that MIFAWN achieves average diagnostic accuracies of 99.38% and 98.06%, respectively, significantly outperforming existing mainstream unsupervised transfer learning methods. Meanwhile, the proposed method exhibits excellent convergence stability and feature separability, which shows great potential in practical application scenarios.