| Li Jin (李 进)*,Lin Zesheng**,Wang Qingguo*,Wang Siyuan**,Zhang Jinjie**.[J].高技术通讯(英文),2026,32(2):133~147 |
|
| HAGNet: a hybrid data-model driven frameworks for machinery fault diagnosis under domain deviation |
| |
| DOI：10. 3772 / j. issn. 1006-6748. 2026. 02. 003 |
| 中文关键词: |
| 英文关键词: multi-sensor fusion, fault diagnosis, domain deviation, hierarchical attention mechanism |
| 基金项目: |
| Author Name | Affiliation | | Li Jin (李 进)* | (* CNOOC EnerTech Equipment Technology Co. Ltd. , Tianjin 300450, P. R. China)
(** Diagnosis and Self-Recovery Engineering Research Center, Beijing University of Chemical Technology, Beijing 100029, P. R. China) | | Lin Zesheng** | | | Wang Qingguo* | | | Wang Siyuan** | | | Zhang Jinjie** | |
|
| Hits: 115 |
| Download times: 129 |
| 中文摘要: |
| |
| 英文摘要: |
| In the field of mechanical equipment fault diagnosis, variations in operational conditions often lead to shifts in data distribution, which consequently results in weakening the diagnostic precision.This study presents a hybrid data-model driven framework for fault diagnosis under domain deviation, termed the hierarchical attention guided generalization network (HAGNet). This method can effectively enhance fault identification accuracy under domain deviation conditions, even when target domain data is unavailable during the model pretraining phase. Firstly, a multi-channel data fusion module (HADF), which integrates multi-head self-attention and spatial attention mechanisms, is introduced. Subsequently, a cross-domain center-alignment and type-separation (CATS) loss function is developed to extract domain-invariant features. Finally, a multi-classifier ensemble mechanism is constructed to address the issue of insufficient classifier capacity. The contribution of each source domain is dynamically weighted according to its similarity to the target domain, mitigating negative transfer effects from dissimilar sources. Experimental results demonstrate that the proposed HAGNet surpasses existing mainstream methods in fault diagnosis performance. |
|
View Full Text
View/Add Comment Download reader |
| Close |
|
|
|
