| 摘 要: 针对滚动轴承故障诊断中单模态表征不完整、跨模态交互不深入及小样本下缺乏置信度的问题,提出自适应多模态跨模态融合网络HBA-MCFNet。该模型构建时域、时频与频域三模态互补特征体系,在交叉注意力融合基础上引入噪声自适应门控与贝叶斯模态加权,并联合中心损失增强类内紧凑度。消融与多噪声实验表明:三模态互补与噪声门控的有效性均获验证；小样本下贝叶斯模型准确率较基线提升4.72个百分点且训练稳定性翻倍；全噪声等级下均优于对比模型。 |
| 关键词: 轴承故障诊断 跨模态融合 贝叶斯推断 不确定性量化 注意力机制 |
|
中图分类号:
文献标识码:
|
| 基金项目: 国家自然科学基金面上项目资助(52475524) |
|
| Bearing Fault Diagnosis Based on Cross-modal Fusion and Bayesian Inference |
|
Wang Ziyuan1, He Xiaoxia1, Li Chunli1, Zhang Liping2
|
1.School of Mathematics and Systems Science, Wuhan University of Science and Technology;2.School of Mechanical Engineering, Wuhan University of Science and Technology
|
| Abstract: To address the issues of incomplete unimodal representation, insufficient cross-modal interaction, and lack of confidence estimation under small-sample conditions in rolling bearing fault diagnosis, an adaptive multi-modal cross-modal fusion network termed HBA-MCFNet is proposed. The model constructs a tri-modal complementary feature architecture comprising time-domain, time-frequency, and frequency-domain representations. Building upon cross-attention fusion, noise-adaptive gating and Bayesian modal weighting are introduced, jointly optimized with center loss to enhance intra-class compactness. Ablation and multi-noise experiments demonstrate the effectiveness of both tri-modal complementarity and noise-adaptive gating. Under small-sample conditions, the Bayesian model improves accuracy by 4.72 percentage points over the deterministic baseline while doubling training stability. The proposed model consistently outperforms the compared models across all noise levels. |
| Keywords: rolling bearing fault diagnosis cross-modal fusion Bayesian inference uncertainty quantification attention mechanism |
