Deep transfer diagnosis based on MTF image and temporal signal fusion with enhanced interpretability for rotating machinery

Abstract

Intelligent fault diagnosis of rotating machinery is often hampered by scarce labelled target-domain data, distribution shifts across sampling frequencies and machines, and the limited transparency of deep models. This paper proposes an unsupervised cross-domain diagnosis framework that fuses one-dimensional vibration signals with two-dimensional Markov transition field images within a deep transfer-learning pipeline. A dual-branch encoder, combining a gated recurrent unit and a ResNet enhanced with a convolutional block attention module, extracts complementary temporal and spatial-texture features that are projected into a shared latent space. To suppress pseudo-label noise during alignment, an improved joint distribution adaptation scheme couples multi-kernel maximum mean discrepancy with correlation alignment and imposes class-conditional, confidence-weighted constraints; an improved Softmax loss further enlarges inter-class margins. A dynamic transfer schedule progressively increases the alignment weight along a sigmoid trajectory, while an adaptive confidence threshold tracks the evolving reliability of target-domain predictions. Class activation mapping visualises the decision basis and provides activation-based evidence related to physical fault mechanisms. The method is evaluated on cross-frequency (12 kHz and 48 kHz) and cross-machine (CWRU, JNU and SEU) tasks, attaining average accuracies of 91.8% and 93.0% on the two task families. Remove-one ablation, per-class precision–recall–F1 metrics, confusion matrices with 95% confidence intervals, and a hyperparameter sensitivity sweep are used to assess the contribution and robustness of the proposed components.