Abstract: To address the problems that fault information of inner-ring cracks and outer-ring wear of belt conveyor idler bearings is not effectively represented, that effective fusion and utilization mechanisms for multimodal information are lacking, and that unstable fusion effects lead to poor generalization ability and low fault discrimination of diagnosis models, a fault diagnosis method for belt conveyor idler bearings integrating multimodal data and Unified Aggregation Visual Transformer (UAViT) was proposed. First, raw vibration signals and audio signals during the operation of belt conveyor idler bearings were collected by a Micro-Electro-Mechanical System (MEMS) optical fiber vibration sensor and a GSD5 intrinsically safe mining acoustic sensor. Second, signal intervals were segmented, and the raw signals were processed using wavelet threshold denoising, Wiener filtering, and short-time Fourier transform to extract time-frequency images of vibration signals and narrowband spectrograms of audio signals, thereby constructing a dataset. Then, a UAViT model containing an Aggregation Knowledge Extraction Module (AKEM) and an Implicit Prior Knowledge Transfer Module (IPKTM) was constructed, and the two modules jointly achieved deep complementary fusion and stable generalized diagnosis of multimodal information. Finally, the UAViT model was trained using a two-stage supervised learning strategy, and fault diagnosis results for the idlers were obtained on the test set. Experiments based on a real-world dataset collected from a mining area showed that the diagnostic accuracy of UAViT for inner-ring cracks and outer-ring wear faults reached 99.67%. Ablation and control-variable experiments verified the effectiveness of Feature Channel Attention Mechanism(FCAM), IPKTM and the denoising method. Comparative experiments showed that the fault diagnosis accuracy of UAViT was 1.27%-6.67% higher than that of existing advanced models. Extreme-noise, atypical-fault generalization, and cross-device transfer experiments systematically revealed the performance boundaries of UAViT under strong noise, unknown faults, and cross-device variation scenarios. The method successfully diagnoses multiple real-world faults in engineering applications, verifying the practicality and reliability of UAViT in the complex underground environment of coal mines.