Abstract
To address the limited representation of compound fault features and the susceptibility of single-modal diagnosis to noise in wheelset–axlebox systems of urban rail vehicles, this study proposes a multimodal diagnostic method integrating vibration signals and infrared thermal images. A wheelset–axlebox fault simulation platform was established to acquire vibration and infrared data in parallel under different equivalent vehicle-speed conditions, followed by temporal pairing to construct a multi-condition dataset containing normal, single-fault, and compound-fault conditions. Vibration signals were transformed into short-time Fourier transform (STFT) time–frequency maps and Gramian angular field (GAF) temporal structure maps and combined with infrared thermal images as three-modal inputs. A Tri-Modal Attention-Enhanced Residual Network (TAER-Net) was developed with three residual branches and a proposed Improved Convolutional Block Attention Module (ICBAM) for feature enhancement and feature-level fusion. The method achieved high classification performance for the tested fault conditions across four speed datasets, with a maximum accuracy of 98.4%. Ablation, comparative, and noise-robustness experiments further demonstrated the effectiveness of the multimodal fusion framework and TAER-Net under the tested conditions. The results provide a methodological reference for multimodal intelligent diagnosis of wheelset–axlebox faults using complementary vibration and thermal-spatial information.