Domain-collaborative multimodal transformer for fault diagnosis of rotating machines under noisy environments

Fault diagnosis in rotating machinery remains a critical challenge due to diverse fault types and vulnerability to signal contamination under severe noise conditions. This study introduces a domain-collaborative multimodal transformer (DCMT), a novel time–frequency fusion framework that jointly analyzes vibration and current signals by preserving domain-specific representations across segmented temporal and spectral pathways. Instead of early-stage fusion, the proposed model processes each modality through independent segmentation, converting signal segments into recurrence plot tensors to capture both local patterns and long-range dependencies. A dual-branch architecture integrates convolutional encoders, bidirectional long short-term memory (LSTM) modules, and transformer layers to extract hierarchical features from each domain. Self-attention modules enable domain-specific representation learning, while a cross-attention mechanism facilitates adaptive fusion between time and frequency branches. Extensive experiments on a real-world subway motor dataset and the Case Western Reserve University bearing dataset demonstrate that the DCMT achieves superior diagnostic accuracy and robustness across a wide range of noise levels. Comparative evaluations with recent benchmark models confirm generalizability and resilience of the proposed framework. An ablation study was conducted to assess the effect of learning-related parameters, architectural configurations, and model components on overall performance. These findings establish the proposed model as a robust and scalable solution for practical fault diagnosis in noisy environments. Source code and pretrained models are publicly available at https://github.com/AVIP-laboratory/Domain_collaborative_multimodal_fault_diagnosis.