Multi-modal evolutionary graph learning-based fault diagnosis for motors under wide-range variable-speed conditions

Motor fault diagnosis under wide-range variable-speed conditions remains challenging due to the dynamic migration of inter-sensor dependencies and the inconsistency of multi-modal feature correlations. Conventional models trained under fixed-speed or single-modality settings fail to adapt to these evolving relationships. To address this, we propose a multi-modal evolutionary graph learning (MMEGL) framework that adaptively captures and aligns heterogeneous feature graphs across varying speeds. A frequency–time adaptive representation learning module emphasizes speed-sensitive harmonics through multi-resolution short-time Fourier transform and learnable Gaussian filtering. Simultaneously, a low-rank adaptation-enhanced w2v-BERT branch is employed to extract domain-adaptive acoustic semantics. A spatiotemporal attention module further decouples and fuses spatial dependencies and temporal dynamics. At the framework level, we establish an evolutionary graph structure learning mechanism based on multi-view mutual-information maximization. This mechanism reconstructs dynamic graph topologies to preserve temporal-correlation consistency across various speed ranges. Experiments on the Ottawa and Korea Advanced Institute of Science and Technology datasets demonstrate that MMEGL achieves 98.22% and 99.27% accuracy, respectively, outperforming 18 advanced baselines. These results demonstrate its robustness under varying signal-to-noise ratio conditions during variable-speed operations.