Trustworthy multimodal feature-enhanced fusion network for non-contact rotating machinery fault diagnosis

Multimodal information fusion and non-contact sensing technology play a crucial role in the fault diagnosis of complex mechanical equipment and have been widely applied in transportation, manufacturing and aerospace industries. However, dynamically assessing the reliability of each modality under low-quality data conditions or significant noise interference remains a major challenge. To tackle this issue, this paper proposes a trustworthy multimodal feature-enhanced fusion network (TMFEFN) framework to enhance the reliability of multimodal fusion learning and improve the extraction of deep, sensitive fault features. Firstly, a dual-branch feature extraction module is proposed to capture both local and global features from acoustics and infrared thermography data. Secondly, an enhanced frequency channel attention network module is designed to refine the unimodal features and construct a combined pseudo-view. Simultaneously, a comparative clustering loss is formulated to enforce consistency among different modal features for each sample in the semantic space. Finally, a trustworthy feature fusion module, based on the Dirichlet distribution, is introduced to measure the contribution of each modality to the diagnostic results, ensuring a reliable fusion of modal features across different samples. The effectiveness and trustworthiness of the proposed TMFEFN method are validated on real-world gearbox and aircraft engine rotor datasets acquired by non-contact sensing technology. Experimental results demonstrate that TMFEFN outperforms five state-of-the-art multimodal fusion methods in both diagnostic accuracy and noise robustness, while also providing a more reliable assessment for the trustworthiness of multimodal fusion diagnostic results.